WO2013122468A1 - Automated monitoring and controlling of undesired livestock behaviour - Google Patents

Abstract

A fully automated method and system for monitoring, controlling or preventing undesired behaviour of an animal or group of animals, particularly livestock animals was developed, making use of the intelligence of the animals and their ability to learn. A trigger or stimulus is automatically generated when automated monitoring shows or predicts the onset of undesired behaviour of an animal. Said trigger or stimulus then prevents or stops the undesired behaviour, particularly by capturing or redirecting the attention of said animal. Preferably, said method and system is also capable of training the animals, particularly livestock, to associate said trigger with a reward, punishment or warning in a fully-automated way. The conditioned animal will then give its attention to the trigger at the moment that early signs of undesired behaviour have been observed or predicted.

Description

AUTOMATED MONITORING AND CONTROLLING OF UNDESIRED LIVESTOCK

BEHAVIOUR

FIELD OF THE INVENTION

The present invention relates to an automated method and system for altering or controlling the behaviour of livestock animals. Particularly, the present invention provides methods and systems for monitoring and preventing and/or correcting undesired livestock behaviour.

BACKGROUND OF THE INVENTION

In livestock production, some animal behaviour, such as aggression and cannibalism pose a serious welfare issue and create great problems in livestock farming management, resulting in economic losses. In the context of the present invention such behaviour is referred to as undesired behaviour.

In pig husbandry, for example, aggressive behaviour (e.g. fighting) or abnormal behaviour (e.g. tail-biting, ear-biting) are undesired behaviours since they cause high stress level: the pigs become more easily subjected to diseases and have lower growth levels. Other undesired behaviour is e.g. cribbing in species such as horses. Cribbing is a symptom of confining a range animal to a small stall, and most horses are confined to small stalls for convenience.

In intensive farming, pigs are kept in a confined environment and express aggressive behaviour on a much higher level than they do in a natural environment. Reasons for this aggressive behaviour in intensive farming conditions can be found in the limited space allowance, barren environment, low fibre feed diets and repeated changes in group composition. In fact, domesticated pigs are hierarchical animals just like wild pigs and in intensive farming, the group hierarchy does not always remain stable due to the commercial practice of mixing the animals. This mixing occurs usually after weaning, at the beginning of the fattening period or in sows after service due to management choice. This practice results in intense aggressive interactions that occur mainly in the period of the first two days from the moment of the new group formation until the new dominance hierarchy has been established. These encounters can lead to wounds that may cause infections and in extreme cases may even be lethal.

Furthermore, aggression leads to economic losses because weaker animals that are dominated by more aggressive animals have not enough access to food, which results in decrease of growth rate and increase of weight variability within the pen. Injuries caused by aggressive interactions can also cause a significant loss in value due to the condemnation of parts of the carcass or the downgrade of the carcass to a lower meat quality. Additionally, stress caused by aggressive behaviour can reduce the fertility of breeding sows. Aggression among pigs is therefore one of the most important health, welfare and economic problems in intensive farming. The cost estimation of tail damage among pigs in the Netherlands indicates a financial loss of over eight million euro per year (Zonderland, J.J. 2010.). This figure includes cost of reduced weight gain, on-farm veterinary treatment, culling and carcass condemnation.

In the case of pigs, different studies have tried to reduce the level of aggressive behaviour in pigs but no suitable method has been discovered so far. The solutions that were found in literature either postpone the aggression (e.g. tranquilizer) or the pigs are treated in a way their welfare is reduced, for example by giving drugs to the animals. Tan and Shackleton administered a commercial tranquillizer Stresnil (Azaperone) to the pig in order to reduce the level of aggression. However, Azaperone eliminated fighting only for a few hours after it was administered, after which the behaviour of the treated pigs was not different from that of non- tranquillized pigs (Tan, S.S.L.; Shackleton, D.M. (1990). In another study, gilts were washed and Vicks was applied to their nostrils in order to mask the odours of the pen mates. The amount of fighting and their intensity was not influenced by the treatment (Luescher, U.A.; Friendship, R.M.; McKeown, D.B. (1990). The effect of the enrichment of the environment by introducing straw in the pens was evaluated. The enrichment of the environment is considered to reduce the aggression since it is considered that one of the reasons for fighting is due to the poor environment, but it was shown it has no effect (Arey, D.S.; Franklin, M.F. (1995):45„ 23-30). An increased tryptophan intake to raise brain serotonin (5-HT) levels and thereby altering the behaviour of pigs to reduce the level of aggression was also evaluated. This treatment decreased the aggressive behaviour especially in 3-month-old gilts but it also decreases the overall activity and the time spent standing, thus decreasing the welfare of the animal (Poletto, R. et al. (2010).

Other documents focused on training methods to encourage good behaviour and/or punish or control the undesired behaviour. US5566645 describes a method for animal training, particularly a method for rapidly and effectively training horses and other animals by facilitating the delivery of a primary reinforcement reward substance to the animal simultaneously with, or immediately following the exhibition of desired behaviour by the animal. GB2473540 describes an animal training apparatus comprising a trigger

arrangement. The apparatus is used as part of a positive reinforcement training programme in which the animal to be trained associates the noise of a clicker with the administration of a fluid reward. US4335682 provides a unit adapted to be worn by a dog or other animal, which acts under the control of a remote control unit to produce stimuli including an aversive electrical stimulus, a characteristic sound or other second stimulus to which the animal responds with a safety, relief and relaxation response, or a warning stimulus. US5351653 describes a method for training an animal based on positive and negative audio signals. The method enables a trainer to encourage good behaviour by the animal by applying the positive audio tone after the animal has been trained to associate the positive audio tone with pleasant feelings and to discourage bad behaviour by the animal by applying the negative audio tone.

However, these documents mainly focus only on pets like dogs or horses and not on livestock. In these methods human interaction/intervention is needed to recognize or detect the unwanted animal behaviour. In this respect, pets live in a relative comfortable

environment, more close to humans from who they receive more monitoring. Livestock animals, instead, are kept in very big groups (e.g. 60.000 broilers). The farmer has no time to train each individual animal. Also, many training methods based on punishment of undesired behaviour include harming the animal in some way. In particular, shocking devices of various kinds are well known in animal training. Such treatment is considered to be cruel.

Thus, there remains a need in the art for animal-friendly methods and systems to monitor, prevent or reduce the level of undesired behaviour in livestock, such as aggressive or unwanted behaviour in pig husbandry, which require little or no human interaction or intervention.

SUMMARY OF THE INVENTION

The present invention provides a fully automated method and system to monitor, correct or prevent the undesired behaviour, by a combination of different triggers or stimuli, which the animal associates with a reward, punishment or warning. Particularly, said trigger or stimuli is initiated when the first or early signs of undesired behaviour have been detected or predicted, thus stopping or preventing the undesired behaviour in a fully automated way. By preventing or minimising the undesired behaviour, the livestock animals experience less stress, thus improving their welfare and their production efficiency.

Automated monitoring of on-farm animal welfare has a number of potential advantages, such as continuous measuring of indicators, real-time registrations, more accurate and precise information and increasing flexibility in the time management of the farmer.

A first aspect of the present invention provides an automated method for controlling or preventing undesired behaviour of an animal or group of animals comprising the steps of automatically monitoring the behaviour of said animal or group of animals and generating a stimulus or trigger to prevent or stop the undesired behaviour. Typically, said automated monitoring detects the early signs of undesired behaviour of said animal or group of animals or generates data that is to predict future undesired behaviour of said animal or group of animals. Once the automated monitoring has detected or predicted the occurrence of undesired behaviour, said stimulus or trigger is generated automatically and prevents or stops the undesired behaviour, such as by capturing the attention of said animal or group of animals. Preferably, said method of the present invention further comprises the step of conditioning or training the animal to associate said trigger or stimulus with a reward, a punishment or warning.

Another aspect of the present invention relates to a system or control device for

automatically controlling or preventing undesired behaviour of an animal or group of animals comprising means for monitoring the behaviour of said animal or groups of animals and means for generating a trigger or stimulus to prevent or stop undesired behaviour. Said means for monitoring the animal behaviour comprises at least one suitable sensor (e.g. a camera, a microphone, a motion sensor, a heat sensor, a heart rate sensor) which collects data on the animal behaviour. Preferably, said system or control device of the present invention further comprises a processing system, capable of processing the data collected by the sensor(s) to detect or predict undesired behaviour, preferably the onset or early signs of said undesired behaviour. Said processing system also commands said means for generating a trigger or stimulus to generate a suitable trigger to prevent or stop undesired behaviour.

Advantageously, the method and control device of the present invention modifies, reduces, prevents or eliminates undesired animal behavior in an automated way, i.e. without human intervention or in the absence of a human supervisor. Preferably, the reduction, prevention or elimination of the undesired behaviour by the animal occurs without punishment of said animal. Particularly, the method and control device of the present invention is used to modify or prevent the undesired behaviour of at least one animal (such as one or more animals showing early signs of undesired behaviour) within a group of animals.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will now be further elucidated by means of non-limiting, examples referring to the drawing in which:

Figure 1 illustrates the general structure of an automated control scheme according to a method of the present invention. Figure 2 illustrates the general structure of an automated control scheme to grab the attention of the animal to control or prevent undesired behaviour

Figure 3 illustrates the different non-contact pre-sign positions before the start of the aggressive behaviour. Codes as in Table 1 a.

Figure 4 illustrates the different contact pre-sign positions before the start of the aggressive behaviour. Codes as in Table 1 b.

Figure 5 shows a Motion History Image of a pig moving from left to right.

Figure 6 shows segmented zones of pig movement from the Motion History Image.

Figure 7 shows the scatter plot depicting the association between the two features extracted from the MHI for both the classes aggression and no aggression. The line is the calculated LDA boundary representing the separation between the two clusters of data.

Figure 8 shows the scatter plot depicting the association between the two features extracted from the MHI for both the classes aggression and no aggression manually selected between episodes with low, medium and high group activity. The line is the calculated LDA boundary representing the separation between the two clusters of data.

Figure 9: The predicted probability of behaviour continuation after sound release.

Figure 10 shows the total number of all behaviours (black) and the number of interrupted behaviours (white) observed at day 1 and day 2 after mixing (five experimental trials).

Figure 1 1 : Reaction of the receiver in relation to the aggressive interactions, which were stopped after feeder activation (from 147 stopped interactions).

Figure 12: Reaction of the aggressor in relation to the aggressive interactions, which were stopped after feeder activation (from 147 stopped interactions).

Figure 13: Type of reaction in relation to the total percentage of encounters.

Figure 14: Type of reaction (none, stopped and not stopped) in relation to the total percentage of encounters.

Detailed Description

As used herein, the term "livestock" refers to any animal or group of animals which is intended to be monitored and/or managed, regardless of whether the animal(s) is

domesticated, semi-domesticated or wild, and regardless of the environment in which the animal may be found, such as, for example, in a commercial animal operation, or in a wild environment. Preferred livestock animals include animal species capable of learning to associate a trigger or stimulus with a reward, punishment and/or warning and include pigs, cattle, goats, sheep, horses, chickens, buffalo, deer or other wild animals. The method and control device of the present invention may also find use for zoo animals or laboratory animals.

In the context of the present invention, a "processing system" includes a system using one or more processors, microcontrollers and/or digital signal processors having the capability of running a "program" which is a set of executable machine code. Processing systems include computers, or "computing devices" of all forms (desktops, laptops, PDAs, servers, workstations, etc.), as well as other processor-based communication and electronic devices such as cell phones, tablets, personal data assistants, etc. Such processing systems may be discrete units, or may be formed of multiple components, which may be networked or otherwise capable of being placed in operative communication with one another, at least at needed intervals. A "program" as used herein, includes user-level applications as well as system -directed applications or daemons.

The inventors developed a fully automated method and system for controlling or preventing undesired behaviour of an animal or group of animals, particularly livestock animals, making use of the intelligence of the animals and their ability to learn. In said method or system a trigger or stimulus is automatically generated when automated monitoring shows or predicts undesired behaviour of an animal. Said trigger or stimulus then prevents or stops the undesired behaviour, particularly by capturing the attention of said animal. Preferably, said method and system is also capable of training the animals, particularly livestock, to associate said trigger or stimulus with a reward, punishment or warning in a fully-automated way. The conditioned animal will then give its attention to the trigger or stimulus at the moment that early signs of undesired have been observed or predicted, thus its undesired behaviour is prevented in a fully automated way.

Thus, a first aspect of the present invention provides an automated method for controlling or preventing undesired behaviour of an animal or group of animals comprising the steps of automatically monitoring the behaviour of said animal or group of animals and generating a stimulus or trigger to prevent or stop the undesired behaviour. Typically, said automated monitoring detects the early signs (before escalation) of undesired behaviour of said animal or group of animals or generates data that is processed to predict future undesired behaviour of said animal or group of animals. Once the automated monitoring has detected or predicted the onset of undesired behaviour, said stimulus or trigger is generated automatically and prevents the escalation towards or stops the undesired behaviour, such as by capturing or redirecting the attention of said animal or group of animals, particularly by the (conditioned) association of said stimulus or trigger to a reward, punishment or warning, preferably a reward.

Although automated systems that detect animal behaviour on a given moment are known in the art (e.g. CN101214168), it is understood that they are often insufficient in the context of the present invention. Indeed, in the context of the present invention it is preferred that the status of the animals is detected in an early stage by recognising early signs of undesired behaviours or even that the future behaviour of the animals can be predicted. Only in an early stage of undesired behaviour, the animal is capable of giving its attention to said trigger(s) and stops the unwanted behaviour.

In a preferred embodiment said method of the present invention further comprises the step of conditioning or training the animal to associate said trigger or stimulus with a reward, a punishment or warning.

The present invention further relates to a system and control device for automatically controlling or preventing undesired behaviour of an animal or group of animals comprising means for monitoring the behaviour of said animal or groups of animals and means for generating a trigger or stimulus to prevent or stop undesired behaviour. Said means for monitoring the animal behaviour comprises at least one suitable sensor which collects data on the animal behaviour. Preferably, said system or control device of the present invention further comprises a processing system, capable of processing the data collected by the sensor(s) to identify, detect or predict undesired behaviour, preferably the onset or early signs of said undesired behaviour. Said processing system also commands said means for generating a trigger or stimulus to generate a suitable trigger to prevent or stop undesired behaviour. Said processing system thus relates the identified or predicted undesired behaviour with the generation of a trigger to stop or prevent said undesired behaviour. Said processing means may also comprise means for storage of data.

In the context of the present invention a suitable trigger or stimulus to capture or redirect the attention of the animal or to stop/prevent the undesired animal behaviour can be an audible, visual or sensory stimulus (e.g. sound, light, vibration, pressure). Particulary, said method or device of the present invention can be used to condition the animal to associate said trigger, such as a particular sound or light, with a reward, a punishment or warning. Any sensory stimulus which is pleasant or unpleasant to the animal may be used as reward or

punishment, respectively, and include food, toy, scent, image, sound, tactile items and combinations thereof. Preferably, in the context of the present invention a reward is used. Examples include, without limitation, food and other edible material dispensed in the form of pellets, toys of the chewable type (chewtoys) made with rubber, plastic, rawhide and the like, puzzle toys and toys which may be filled with food, a brush, scents dispensed in the form of a spray, a breeze using a tunnel fan, vibrations using a vibrator and/or rocker pad, visual items and recordings.

In the context of the present invention, the automated monitoring makes use of at least one suitable sensor. Suitable sensors include at least one or more of the following:

- a camera;

- a microphone, which may be contained in the housing or mounted to a collar on the animal; - a motion sensor to detect e.g. a specific motion pattern (pacing);

- a heat sensor for sensing body heat of the animal, particularly livestock animal;

- a heart rate sensor for sensing the heart rate of the animal, particularly livestock animal;

For example, chewtoys may be equipped with pressure and/or tension sensors to monitor chewing behaviour. A pressure-sensitive pad may detect a position of the animal thereon, which may correspond to a desired behaviour such as resting, or absence of pacing, door scratching and the like spatially fixated undesired behaviours at other locations.

A particular embodiment of the control device of the invention may include a processing system operatively executing one or more algorithms or programs for analyzing inputs received from the at least one sensor, said one or more algorithms being configured (A) (i) to identify the inputs the occurrence of undesired behaviour, particularly the onset or early signs of such undesired behaviour; or (ii) to predict the future occurrence of such undesired behaviour; and once such undesired behaviour has been detected or predicted (B) to command the generation of a trigger or stimulus to stop such undesired behaviour.

Preferably, in the context of the present invention, animal behaviour is monitored in a continuous way with real-time registration and processing, with an accuracy that allows the continuous detection and/or prediction of specific behavioural aspects.

Additionally, the automated monitoring may include collecting data that will be useful in identifying individual livestock. This may be implemented with a tag, particularly a tag that is machine-readable only. A tag relates to any device allowing identification of an individual animal, regardless of how the device may be associated with an animal, such as by being externally affixed to the animal (for example, in the manner of conventionally-known ear tags, by a collar, or by some other mechanism), or by being implanted or otherwise internally carried by the animal. For instance, said tag may be a passive, machine-readable radio frequency identification ("RFID") tag associated with an individual animal.

A method for controlling or preventing undesired behaviour in a fully automated way is schematically shown in Figure 1 . The animal or each individual animal in a group of animals is monitored continuously by using suitable sensors, such as cameras, microphones, etc. These sensors are used to extract variables from the animals. These variables are used to detect the early signs that will lead to undesired behaviours or are used to predict future undesired behaviour. In order to control or prevent the undesired animal behaviour their attention is captured by using one or more triggers or stimuli and the animal will not enter into a specific stage or specific mental status of undesired behaviour (e.g. fighting, biting, ...) (Figure 2). If the attention of the livestock animal is obtained, then the animal can be preferentially rewarded (e.g. giving feed, ...). Alternatively, it can be "punished".

Advantageously, the undesired behaviour is detected or predicted as early as possible. Once the animal enters too far in a specific stage or specific mental status of undesired behaviour it is difficult, or even impossible, to grab their attention and control their behaviour.

Preferably, the reward/punishment, more preferably reward, or warning is used as part of a reinforcement learning in which the livestock is conditioned to associate the trigger(s) or stimuli (e.g. light or sound) with the application of a reward/punishment. The distribution of punishment and reward is adjusted automatically in the time of the training. At the beginning of the training the rewards/punishments will be continuously given in response of the attention given, but it will decrease over time and at the end the reward/punishment will be only given in order to keep and reinforce the association between reward/punishment and the trigger(s).

The schematic diagram presented in Figure 2 shows how in the present method and control device the attention of the animal can be captured. The animal is the central process of the system and it is affected by the environment and by the social interaction with the other animals. In order to grab the attention of the animal, trigger(s) are used. Possible triggers are sound, light, smell, etc. The triggers are changed and adjusted depending on the response of the animal in a close loop control system.

An automatic animal behaviour control device or method of the present invention may reduce or eliminate the frequency or occurrence of any undesired behaviour that increases the stress level in livestock animals, including but not limited to aggressive behaviour, self- mutilation, biting, etc, that is detectable and that can be detected and processed as an electronic signal. Because the attention of the animal is captured or diverted and/or the animal is likely to settle down calmly, the automated method and device of the present invention will also reduce the level of stress and, consequently, increase livestock welfare and productivity.

Example 1. Early phase or pre-signs of aggression interactions

Study setup

The experiment was carried out at a commercial farm, located in Heusden (the Netherlands), with a capacity for approximately 6000 fattening pigs, weighing from 23 - 120 kg. The farm utilizes ad libitum dry feeding systems (Fancom B.V. - F71 ) and a central flow ventilation system (Fancom B.V. - F21 ). Behavioural observations were carried out on a newly composed group of 1 1 entire male pigs of 23 kg on average kept in a pen of 4m x 2,5m with partially slatted concrete floor and solid pen walls. Pigs were sprayed on their backs with standard colour stock marker to facilitate the identification of individuals in order to identify their behaviour before the start of aggressive interaction.

Video recordings were performed using a camera (Allied Vision Technologies®, model F080C), placed above the pen in central position (in top view), at the height of 2.3m, that permitted an overview of the whole pen. The camera was connected to a computer installed inside the room. Strong paper wall was protecting pigs from disturbances associated with computer and human presence in the room. Images were captured with a frame rate of 1 1 frames per second, resolution of 1032 x 778 pixels, in colour. A total of 8 hours of video recordings were registered during 3 days after mixing. The video recordings were carefully observed to detect aggressive interactions between pigs.

The aggressive interaction was defined as a close physical contact, during which at least one of interacting pigs performed head knocking, biting or pressing behaviours deemed to have ended after retreat of one or both pigs with separation lasting at least 3 sec. If after the separation one of the pigs attacked immediately the other animal it was counted as new interaction. Each interaction was observed (labelled) on the video images frame by frame (1 1 frames per second) to register its exact duration and to describe the behaviour and body positions of pigs on the early phase of aggression. The following initial behaviours of initiator pig at the first moment of aggressive interaction were labelled:

- Body biting: assigned when initiator started aggressive interaction with biting (opened its mouth and closed it on another pig) of any part of the body of another pig, excluding the front third of the body (head, ear, neck). - Head biting: assigned when initiator started aggressive interaction with biting at the head region (except ears) of another pig.

- Neck biting: assigned when initiator started aggressive interaction with biting at the neck zone and shoulders of another pig.

- Ear biting: assigned when initiator pigs started aggressive interaction with biting at the ear of another pig.

- Head knocking: assigned when the initiator pig uses a vigorous fast side to side or upwards movement of its head to hit any part of the head or body of another pig. The mouth is kept closed.

- Mount: assigned when the initiator pig starts aggressive action by jumping on responder pig with its forelegs from lateral side or rear.

In order to describe an early phase of aggression interactions and to identify their possible pre-signs, the 13 body positions of two interacting pigs before the start and at the start of an aggressive interaction were labelled. They were classified into 5 non-contact (on distance) pre-sign positions leading to aggressive interaction and 8 contact body positions (Table 1 , Figures 3 & 4).

A non-contact pre-sign position was defined as a body position of an initiator pig towards another pig before the start of aggressive interaction. The duration of non-contact pre-sign position was registered from the moment when the initiator pig was noticed to raise its head in the direction of the other pig prior to approaching till the start of aggressive interaction.

The contact body positions were defined as the body positions of the pigs already approached to each other and resting in close contact by any part of their body (Table 1 , Figure 4). They were labelled as contact pre-sign positions when the animals were staying in these positions for at least 1 second before the start of aggressive interaction, while at the first moment of the start of aggressive interaction the same positions were registered as contact body positions at the start of aggressive interaction. Duration of contact position as pre-sign was recorded from the moment of the first contact of any part of the pigs' body till the start of aggressive interaction. The duration of aggressive interaction was registered from the moment of the detection of contact aggressive interaction start position till separation of the pigs. If aggression started immediately, without pre-sign position noticed, it was labelled as an aggressive interaction without pre-sign (pre-sign position noticed on the video less than 1 sec before the start of the aggressive interaction was not considered). In this case for this aggressive interaction only contact aggressive interaction position was registered. Table 1 a: Description of labelled non-contact pre-sign body positions

Statistical analysis: Data was processed using Frequency procedure (Prog. Freq SAS, 201 1 ), Chi square test for enumeration data, to evaluate the type of each pre-sign in relation to the aggressive episodes. A total of 177 aggressive interactions were identified from 8 hours of video recordings. The duration of most of registered aggressive interactions (41 .24%) were short, from 1 to 5 sec (see Table 2).

Table 2: Duration of aggressive interactions

The non-contact pre-sign positions could be noticed before 55% of observed aggressive interactions (Table 3), while the contact pre-sign positions were observed only in 15 % of the cases. Most common pre-sign positions were P3 (43 pre-signs=24%, P>0.001 ), when pigs approached facing each other and P2 (32 pre-signs=18%), when the attacking pig approached from the lateral side. Aggressive interactions most commonly began with the animals in the P12 inverse parallel position (39.5% of all bouts), P7 position nose-to-nose forming 90 ° angle (19.7%) or P9 in perpendicular position with nose approaching to anterior part of the body (13.5%).

Table 3. Number and percentage (%) of observed positions

The video data showed that aggressive pigs move towards their opponent with high speed. Within 1 -2 seconds the attacking pig, starting in a non-contact pre-sign position, bridged the distance to the opponent (Table 4).

The contact pre-sign positions in some cases lasted more than 2 sec, when one of the pigs was following another one while not breaking body contact or in the case of mounting. Table 4. Duration of pre-sign positions (sec)

The most frequent combination of pre-sign and aggressive interaction start position was PS- PI 2 (Table 5). P7 and P12 interactions without pre-signs were observed relatively often as well (P7 without pre-sign=1 1 interactions; P12 without pre-sign=18 interactions). As for all the interactions without pre-signs the pigs were on short distance to each other, so the start of attack could be noticed on the video in less than 1 sec before the start of aggressive interaction. Table 5 Relation of pre-sign positions to contact body positions at the start of aggressive interaction.

Table 6 shows the initial behaviour of the initiator pig, with which it started an aggressive interaction and the contact body position from which it was performed. Among the aggressive interaction initial behaviours the most frequent was the head knocking (34.46 %). Mostly, this initial behaviour was observed in relation to P12 and P7 positions. The bites were particularly directed to the neck and ears. From P12 position pigs started aggressive interaction with biting more frequently than from other positions (Table 6), mostly directed towards the neck. Sometimes pigs bit at other regions of the body, particularly flanks or back (6.8%). In 34% of all aggressive actions, no particular behaviours considered for the labelling in the current study were observed at the start of aggressive interaction. Mounting was a pre-sign only for 3 aggressive interactions, in most of the cases this behaviour of initiator pig did not escalate towards an aggressive response from the receiver. Table 6. Initial behaviour in relation to the positions of the pigs' body at the start of aggressive interaction

HK-head knocking; HB-head biting; NB-neck biting; BB-body biting; M-mounting

The analysis of the labelling of the aggressive behaviour at pre-sign and start positions is a first step in the development of an automatic detection of aggressive behaviour in pigs. These results allow identifying key indicators of pig aggression in the early stage of aggressive interactions which can form the basis of an image-based aggression monitoring system. As far as we know there is no other study which investigated the early signs of aggressive behaviour in detail.

Most literature data describe the typical fighting positions in pigs when fighting has already started. Jensen (1980) described aggressive interactions as sequences of different behaviours (parallel/inverse parallel pressing, head-to head knock, head-tilt etc.) and positions (nose-to-nose, when nose approaches the snout or the head of the opponent, nose-to-body and anal-genital nosing). McGlone (1985) described behavioural sequences related to aggression from the first bite to the first sign of submission. The description included different categories of behaviours, such as orienting (pigs facing each other, pigs side-by-side, etc.), bite targets (head, ear, neck/shoulder, posterior parts), pushes, fighting strategy (as jump with front feet on others head/neck area). The spatial orientation (positions) that pigs adopt during the fight were described by Rushen and Pajor (1987). In their study of offence and defence mechanisms of fights they showed five spatial configurations: "T" - configuration, head-to-head, reverse parallel, asymmetric parallel and parallel.

D'Eath and Pickup (2002) adapted for their aggressive interaction ethogram fighting postures described by McGlone (1985) and Rushen and Pajor (1987). They have reported that most of attacks started from T-position-head, head-to-head and direct parallel. In our case P12 position (head-to-head) was the most frequent aggressive interaction start position (39.55%), which is similar to results of Rushen and Pajor, 1987 (37 % of all bouts), while D'Eath and Pickup showed that most of attacks occurred from T-position-head (similar to P9 and P7). The reason of the start of aggressive interactions from these positions could be understood from explanation of fighting mechanism given by Rushen, who showed that fighting pigs spent most of time in the head-to-head configuration, biting each other at a similar rate. However, as bites are delivered most effectively from the lateral position, pigs in order to move from head-to-head position, push against each other snouts and shoulders (inverse parallel pressing) as an attempt to lever their opponents around. Thus, head-to head position could turn into T-position-head when the responder turned away through 90 ° (D'Eath and Pickup, 2002). Numerous studies of aggressive behaviours showed that ears,

neck/shoulders and head are the main target zone for bites during the fights (eg.

McGlone, 1985, Rushen and Pajor, 1987). These studies agree with these results, since at the start of the aggressive interactions the bites were directed mostly to the neck and ears. From P12 position pigs started aggressive interaction with biting more frequently than from other positions, mostly the neck, as this target zone was the most achievable for the bites from this position. However, the most frequent initial behaviour of aggressive interaction was the head knocks, mostly from P12 and P7 positions. Most of aggressive interactions (72%) were short from 1 to 10 sec. These behaviours are supposed to be a short demonstration of dominance when the hierarchy is already established. Rushen (1987) stated in his study that the fights should be brief, when the opponents already know the fighting ability of each other after prolonged fights, in this case one animal will threaten or attack, and the other will

immediately retreat.

We found that not only body position at the start of aggressive interactions but also positions from distance (without any contact) in 55 % of cases relevant to later aggressive behaviour. Most frequent distance positions were P3, when an initiator pig arrived directly facing another pig and P2, when a pig approached the opponent from the lateral side. Arriving from P3 pre- sign position, most often pigs started interaction from P12 inverse parallel position. From contact pre-sign positions pigs started aggressive interactions just in 15%. In 28% of aggressive interactions no pre-sign positions were identified.

In conclusion, most of the aggressive interaction pre-sign positions could be identified in the video, allowing identifying aggression before fighting starts. The early signs and body positions found in this study can be further developed in an image-based monitoring system, able to identify aggression in an automated way before it starts or in a very initial stage before it escalates into an injuring fight. Coupling this image-based monitoring system to a trigger/stimulus generating system (with the pig conditioned to associate the trigger/stimulus to a reward, punishment or warning) allows intervening automatically before the aggressive interaction starts. This way, the problem of excessive aggression among pigs on farms can be reduced. This reduction of aggressive interaction is important for the improvement of pig welfare. Turner et al. (2006) showed that the duration of the fight contribute to a number of lesions, thus intervention before or at the start of aggressive interactions would prevent the suffering of animals from prolonged pain from lesions and wounds caused by long fights. Also, the European Union banned the use of sow confinement stalls by 2013, which means sows should be housed in social groups, where mixing occurs. This automated monitoring and control system of pig aggressive/undesired behaviour will also reduce the economic losses to the producer related to problems with sow fertility, piglet production and daily weight gain.

Example 2. Aggression monitor

Study Setup

ANIMALS & HOUSING. Five repetitive experiments were conducted at the Ruthe

experimental farm of the Hannover Institute for Animal Hygiene (TiHo). In each of the experiments, a total of 24 piglets of the German National Breeding Programme (BHZP) were selected from four litters of piglets (Λ/=120). From birth until weaning, the piglets were housed with their littermates and dam in a 2 m x 2.3 m pen with partly slatted floor and equipped with a farrowing crate, heated piglet area and water and dry feed ad libitum. At the age of five weeks, the experimental animals were weaned and mixed together. From each of the four litters, six piglets weighing at least 10 kg were randomly selected for the experiment and mixed into two pens (N=24). The pens had a dimension of 2 m x 1 .8 m and were equipped with slatted floor and solid pen walls. The piglets had ad libitum access to dry feed and water and the animal feeding place ratio was 1 .5:1 .

The experimental phase started after mixing and lasted 2 days until a new hierarchy among the animals was formed. VIDEO RECORDING. Video recordings of this mixing phase provided a dataset that was used to classify aggressive interactions among piglets.

Video were captured for the first 3 hours after the groups were established and then for 3 hours at approximately 24 h post-grouping. The idea behind was that during the first 3 hours after mixing the pigs have the most severe fights. A relatively short time was needed because of the time consuming labelling procedure since the videos are observed image by image (25 images per sec) to detect all aggressive acts.

A total of 60 hours (3 hours per day) of videos were recorded by two cameras placed 2.0 m central above each pen in order to have a top view perspective. The first camera was a Guppy F-080C camera (Allied Vision Technologies, Germany). The camera used a SV-

03514 3.5 mm lens (VS Technology, Tokyo, Japan). It recorded at a resolution of 1032 ^χ 778 pixels. The second camera was a Guppy GC1350 camera (Allied Vision Technologies, Germany). The camera used a Pentax 4.8 mm lens (Pentax Corporation, Tokyo, Japan). It recorded at a resolution of 1360 ^χ 1024 pixels.

Both cameras were connected to a computer with LabVIEW (8.6, National Instrument, TX) that recorded synchronised videos in MJPEG. The computer's processor was Intel(R) Core(TM) 2 Quad CPU Q9300 @ 2.50GHz with 6 GB of physical memory. The operating system was Microsoft Windows 7 Ultimate.

DATA LABELLING. From the 60 hours of recorded videos, a total of 228 episodes of aggressive interactions were identified and manually labelled by an expert observer who used a software tool developed in Matlab for that purpose (R2009a, The MathWorks Inc., MA).

As mentioned, an aggressive interaction was defined as a close physical contact which lasted at least five seconds and in which at least one of the interacting pigs performed head knocking, biting, or pressing behaviour. When an aggressive interaction stopped, for example due to the retreat of one or both pigs, this sequence was interpreted as finished and any further interaction was considered a new episode. The starting and ending time of every interaction was therefore determined and used as a reference for the classifier.

The labelling procedure is necessary in supervised learning in order to infer an unknown probabilistic function P(x, t) between inputs x e X and labels t e L. This function is called classifier when the output is discrete. The classifier can only be inferred from labelled data {(x„ ti) I i = 1, . . . , nj, where x„ t, are drawn independently from P (x, t). DATASET. In order to evaluate the algorithm, a dataset of 150 episodes with and 150 episodes without aggressive interactions was built (Table 7).

The 150 episodes with aggressive interactions were randomly selected from the 228 episodes manually labelled by the expert.

The 150 episodes without aggressive interactions were built in two steps: 100 episodes without aggressive interactions were randomly selected, while 50 episodes were manually selected by the expert from 1 1 episodes with low group activity (up to 50% of pig moving, up to 50% of pigs resting), 25 episodes with medium group activity (50-80% of pigs moving) and 14 episodes with high group activity (80-100% of pigs moving). This manually selected data was used as a validation of the algorithm in order to prevent that the randomly selected data without aggressive interaction were generated from instances without any activity (i.e. during sleeping).

Table 7. Dataset used for classifying aggressive interactions. The dataset consisted of 150 episodes with aggressive interactions (randomly selected) and 150 episodes without aggressive interactions (100 randomly selected and 50 manually selected between episode with low, medium and high group activity). In the table is reported the minimum, maximum, mean and standard deviation of the duration of each category of episodes.

. Low Activity; b. Medium Activity; c. High Activity Algorithm

The Motion History Image (MHI) is a static image that represents how motion is moving by describing the pixel intensity as a function of the motion history at that point. The result is a scalar-valued image (Figure 5) where brighter values correspond to more recent motion. To generate the MHI for movement, the successive image differences / were weighted and layered. A threshold τ (τ = 1 second) was used to set the time window of the duration for which the motion information was kept: and MHI(x, y) < (frame

For this study, the MHI was implemented in Matlab (R2010a, The MathWorks Inc., MA). The values of the MHI were rescaled between 0 and 255 pixels in order to obtain a grey scale image. This grey scale image was segmented in order to extract local regions of motion (Figure 6). Since the aggressive interactions happened between at least two pigs and since the mean pixel size of one pig is 20000 pixels, the segmented zones of movement smaller than 24000 pixels were filtered out and excluded from further analysis.

PARAMETERS EXTRACTION. Describing the temporal and local motion through the MHI and detecting local regions of movement was an important step to understand image data, but it could not classify whether there were aggressive interactions or not. Another step is therefore to generate a numeric feature vector that characterised the properties of the movement. Two different features were extracted from the segmented regions and their means were used for the evaluation of each episode.

Featurel is a scalar specifying the mean of all the intensity values in the region. This feature represents how strong and intense the motion in the image is.

Feature2 is a scalar representing the occupation of the movement inside the regions and is calculated by the ratio of pixels unequal to zero in the region and the total number of pixels in the region. This feature thus gives spatial information about the movement.

Linear Discriminant Analysis (LDA) is a method to find a linear combination of features that separate two or more classes. A discriminant function L is a latent variable, which is defined as a linear combination of independent variables, where w, are the discriminant coefficients, x, the discriminant variables and k a constant. L = k + Wj j

In LDA, the discriminant coefficients w, maximise the distance between the means of the dependent variables. In a binary classification, only one discriminant function is needed to classify whether the episode belongs to one class {L>0) or the other (L<=0).

In this study, the LDA was used to classify if there were aggressive interactions in the video episodes, using featurel and feature2 extracted from image processing.

Data analysis

SPSS (20, IBM, NY) was used for the LDA to classify aggressive and not aggressive interactions.

The first step consisted in the calculation of the discriminant coefficient of the LDA function based on the features extracted from the MHI.

Afterwards the confusion matrix was calculated. The confusion matrix is a matrix in which the rows are the classes defined by the expert and the columns are the predicted classes. From this matrix, statistical measures of performance such as sensitivity, specificity and accuracy were retrieved.

Sensitivity measures the proportion of actual positive values which are correctly classified. Specificity measures the proportion of negative values which are correctly classified.

Accuracy measures the proportion of the total instance correctly classified.

In order to have more reliable results of the classifier, the data were also cross-validated by using the leave-one-out method. The leave-one-out method uses a single observation from the original sample as validation data and applies the remaining observations as training data. This method is repeated until each observation in the sample has been used once as validation data.

Some of the 300 episodes did not present a feature value because either no motion or too little motion was present and the episode was therefore filtered out. In order to take these values into account during the final performance measurement of the classifier, the final step consisted in adding the true negative (cases that were filtered out and showed no aggressive interaction), the false negative (cases that were filtered out but showed aggressive interaction) and in finally recalculating the confusion matrix for both cross-validated and not cross-validated data. RESULTS

The episodes that did not provide any feature information because they were filtered out due to no or only little movement in the MHI nevertheless contributed to the final result of the classifier. Table 8 shows that 2 episodes with aggressive interaction were filtered out and therefore counted as false negative, while 31 episodes without aggressive interactions were filtered out and therefore counted as true negative.

On the remaining 267 out of 300 episodes, a discriminant analysis was conducted to classify whether there were aggressive interactions or not in a video episode. Predictor variables were the two features extracted from the MHI, namely, the mean intensity {featurel) and the occupation index {feature2).

Table 8. Episodes filtered out and excluded from further analysis. These episodes were taken into account during the evaluation of the classifier's general performance as 2 false negative and 31 true negative.

a. LA: Low Activity; b. MA: Medium Activity; c. HA: High Activity.

Table 9 illustrates the mean differences and standard deviation between the two features in the two different classes. Table 9. Descriptive statistical information of the two features used to classify aggressive interactions.

From the discriminant coefficients, the discriminant function was calculated (Table 10):

L = 2.59 * featurel + 7.603 * featurel— 6.215

Table 10. Discriminant coefficient obtained by using the Linear Discriminant Analysis.

In Figure 7 and Figure 8, the scatter plots show the two clusters of episodes with and without aggressive interactions and the calculated LDA boundary.

Table 1 1 illustrates the results of the LDA classifier, using leave-one-out cross-validation. As can be seen from the confusion matrix, 133 episodes with aggressive interactions and 108 episodes without aggressive interactions were correctly classified. These results indicate an accuracy of 88.4%, a sensitivity of 89.9% and a specificity of 86.7%.

Table 1 1 . Confusion matrix of the Linear Discriminant Classifier for both the original and the leave-one-out cross-validation dataset, without the episodes that were filtered out.

a. 89.9% of original grouped cases correctly classified.

b. 88.4% of cross-validated grouped cases correctly classified.

When the results of episodes filtered out as true and false negative were added, the accuracy becomes 89%, the sensitivity 88.7% and the specificity 89.3% (Table 12). Table 12. Confusion matrix of the Linear Discriminant Classifier for both the original and the leave-one-out cross-validation dataset, including the episodes that were filtered out.

a. 90.3% of original grouped cases correctly classified.

b. 89.0% of cross-validated grouped cases correctly classified.

The above results clearly show that local temporal and spatial information about dynamic regions segmented from the Motion History Image can be used to extract different features related to aggressive and undesired behaviour among pigs. Both extracted features could be used for classification. However, the occupation index provided more information about the data variance and therefore had a higher weight (7.603) in the discriminant function compared to the mean intensity (2.59). This can be explained by the fact that intense motion does not necessarily result from aggressive behaviour, but might as well be caused by other behaviour such as chasing or playing. In order to improve the results and to reduce the false positives, little motion or motion involving only one pig were excluded from this analysis.

Image processing has been used to calculate information about the pigs' activity by means of the activity index in the study of Costa et al. (2009). However, in the Costa study the activity information was extracted from an entire pen or from fixed zones within a pen, while in the present study the use of the MHI provides both spatial and temporal information that is calculated automatically from the motion of the animals and is thus not bound to predefined zones. In addition, the most crucial disadvantage of using the activity and occupation index as in the Costa study (2009) is the fact that movements caused by different kinds of behaviour were summed up and could not be discriminated when occurring within the same zone. Stated differently, the activity and occupation index provides temporal information only, but no spatial information over time. By using the MHI, however, no fixed zones needed to be defined. Instead, zones were calculated dynamically, by using the segmented regions of motion, and analysed separately. Therefore, the method exploited in this study provided more valuable information to detect aggressive interactions among pigs.

ADVANTAGES

According to the Welfare Quality® protocol, the farmers should check the health and welfare status of their animals by assessing injuries in the pen that indicate occurrences of aggression. As this procedure is time consuming and labour intensive, an automatic aggression monitoring system would be beneficial to both farmer and animal.

The present monitoring tool that can continuously and automatically detect aggressive behaviour and consequently monitor the level of aggression in each pen is therefore a valuable tool and can be used by the farmer to increase the animals' health and welfare and to decrease the economic losses. With accurate information about the aggression level in each pen, the farmer can intervene more quickly by separating aggressive animals or by introducing environmental enrichment material in order to reduce the aggression level. It may be argued that aggression levels often return to the same level after a certain period of time due to habituation. However, by continuously monitoring the level of aggression, the environmental enrichment could be changed in order to prevent the effect of habituation whenever the level of aggression exceeds a certain level. As a result, growth rates and uniformity of pigs as well as fertility of breeding sows could be improved. In addition, the present approach does not involve high costs and does not interfere with the animals.

Alternatively and preferably, this monitoring tool able to identify aggression in an automated way before it starts or in a very initial stage before it escalates into an injuring fight can be coupled to a trigger/stimulus generating system (with the pig conditioned to associate the trigger/stimulus to a reward, punishment or warning) which allows to intervene automatically before the aggressive interaction starts. In conclusion, in this example, a method based on Motion History Image was used to calculate dynamic local temporal and spatial information about the mean activity and occupation index in order to detect aggressive interactions among pigs. The results revealed a classification accuracy of 89%, a sensitivity of 88.7% and a specificity of 89.3% and proved that the two motion features (occupation index & mean activity) can be successfully used in order to discriminate between aggressive and nonaggressive interaction.

The aim of this study was to develop a method to automatically detect aggressive behaviour among pigs by means of image processing. 24 piglets were mixed in 2 pens after weaning and captured on video for a total of 60 hours in 5 repetitive experiments. From these video recordings, a dataset containing 150 episodes with and 150 episodes without aggressive interactions was built through manual labelling. The Motion History Image was used to gain information about the pigs' motion and to relate this information to aggressive interactions. Two features were extracted from the segmented region of the Motion History Image, namely, the mean intensity of motion and the occupation index. Based on these two features, the Linear Discriminant Analysis was used to classify the presence of aggressive interactions in every episode. Applying leave-one-out cross-validation, the accuracy of the system was 89% with a sensitivity of 88.7% and a specificity of 89.3%.

Example 3. Reduction of the frequency and duration of aggressive behaviour

Study Setup

In recent years, cognitive abilities of animals were widely tested, showing that pigs can learn successfully to cope with difficult experimental tasks. For the training the classical

(Pavlovian) conditioning can be used to create an association between the food reward and the sound. The animals were taught by operant conditioning to recognize an individual sound and discriminate the sound from other sounds. These previous experiments showed that sound and feed are effective stimuli for the instrumental learning in pigs, and that the pigs can clearly and selectively successfully associate between the sound and the feed reward.

In the present study the associative instrumental learning based on classical conditioning techniques was used as an approach to reduce the incidence of aggressive and abnormal behaviours of pigs reared in intensive conditions. For this purpose we used a prototype of a food-rewarding device for cognitive enrichment, represented by an automatic dog feeder. The piglets learned to approach the feeder which released some attractive feed after hearing the sound signal.

The enrichment tool consisted of a commercially available electronic dog feeder (Manners Minder Treat and Train®, Sommerville Ct.,USA) filled with potentially attractive feed for piglets (chocolate candies). The feeder was activated by remote control releasing a sound signal just before feed distribution.

The study was conducted at the research farm Ruthe of the University of Veterinary Medicine Hannover, Foundation, Germany.

An experimental trial (N=5) consisted of two consecutive phases. Phase 1 (training phase), lasted eight days and aimed to teach the piglets to recognise the association between the sound and the feed reward representing a classical conditioning paradigm (Angermeier, 1994).

The training was carried out on four litters of 25 days old piglets (the German National

Breeding Programme (BHZP)) with average weight of 7 kg (±1 ). The piglets had been raised from birth until weaning with their littermates and dam in a pen (1 .80 x 1 .80 m) with partly slatted floor, equipped with farrowing crate, heated piglet area and provided with water and dry feed ad libitum. Two dog feeders were positioned on opposite walls of the selected pens on height of 0.8 m from the ground and contemporary activated by an observer from outside of the room in order not to distract the piglets by human presence. The sound was played by the feeder and 2 s later the candies were dispensed. During the 2h training period the dog feeders were activated every ten minutes. During each day the training round was repeated five times.

Phase 2 (mixing phase), lasted two days and aimed to test the response of animals on the feeder sound during the performance of aggressive and abnormal behaviours. On the day of weaning from each trained litter 6 piglets were selected based on their weight (average 10kg ± 1 ) and mixed in two pens, 12 piglets per pen. The dimension of the pens was 2m x 1 .8m (3.6m³) with slatted floor and solid pen walls. The piglets had ad libitum access to dry food and water and the animal to feeding place ratio was 1 .5 to 1 . Direct observations were made for the first 3 h after the groups were established (day 1 ) and then for 3 h approximately 24 h post-grouping (day 2). Simultaneous observations were carried out by two observers, one assigned to each experimental pen. The observers were separated from the piglets by a wooden wall in the front of the pens with a small window. The dog feeder was placed on the lateral wall of the experimental pen and distantly activated by the observer when the behaviour described in Table 13 was noticed.

The experimental trial lasted for 2 weeks up to the 6^th week of age. After the experiment, the pigs were finished following standard fattening production until slaughter weight Table 13: Description of behaviours of the piglets leading to the activation of the feeder. Aggressive Behaviour

Video recordings and data analysis

The experimental phases were recorded by two video cameras, Guppy F-080C and. Guppy GC1350 (Allied Vision Technologies, Germany) placed at the height of 2.0 m above the pen floor. Both cameras were connected to a computer with LabVIEW (8.6, National Instrument, TX) that recorded synchronised videos in MJPEG format with variable image rates between 10 and 20 images per second, resolution of 1032 x 778 pixels for the F080C camera and 1360x1024 for the GC1350 and both in colour. The computer's processor was Intel(R) Core (TM) 2 Quad CPU Q9300 @ 2.50GHz with 6 GB of physical memory. The operating system was Microsoft Windows 7 Ultimate.

The recorded videos were analyzed and manually labelled by one observer using a software tool developed in Matlab for that purpose (R2009a, The MathWorks Inc., MA, USA).

Phase 1 data analysis

The database of the Phase 1 (56 h) was analysed to estimate the learning performance of the piglets per each day. The number of piglets around the dog feeder 5 s after the sound exposure was counted.

Phase 2 data analysis

The database of the Phase 2 (54 h) was analysed for the following parameters:

- The exact duration of each behavioural event (the start and finish time of an

aggressive or abnormal behaviour)

- The behaviour of the piglets at the moment of feeder sound exposure (Table 1 )

- The response of the pig on the feeder sound during the performance of the behaviour (0=continued behaviour; 1 interrupted behaviour, approached the feeder)

Feeder latency: latency of response to the feeder sound and interruption of behaviour

The labelling procedure permitted the identification of every selected behaviour happened during a certain period of time. Each recorded video was visually checked image by image (25 images per second) when an aggressive or abnormal behaviour was noticed on the video.

Statistics

Data analysis included five trials. The results of the Phase 1 are expressed as percent of piglets around the feeder 5 s after the sound exposure. The changes in learning behaviour were analysed by a repeated-measures analysis of variance (GLM procedure; SAS, 2008) with the trial as fixed factor and the training day as fixed and repeated factor.

A least squares analysis was carried out on duration of behaviours and feeder latency. Due to the small number of events, push rooting disk, lifting other and tail biting behaviours were excluded from analysis. For the first dependent variable the model included the fixed effects of the day (1 , 2), the response on the feeder sound (1 , 2), the behaviour (1 ,6) and the interaction between the response on feeder and behaviour , whereas for feeder latency the model included the fixed effects of trial (1 , 5) and pen (1 , 2). The comparison among the least square means was carried out by t-test. For these analyses the GLM procedure was used (SAS, 2008).

Logistic regression was used to model the effect of duration as covariate on distraction. The estimated regression coefficient of the duration on distraction was used to estimate the predicted values of logits. A second model was fitted to evaluate the effect of behaviours on distraction. Parameter estimates were obtained from the CATMOD procedure of SAS (2008). Odds ratios, 95% confidence intervals and predicted values of logits were calculated according to the methods of Hosmer and Lemeshow (1989). The odds ratio is a measure of how much more likely (or unlikely) the outcome is among observations with a given risk factor, compared to those without the risk factor. The 95 % confidence interval an odds ratio implies that the true parameter value lies between the two end points 95% of the time (Kleinbaum et al.. 1982). A risk is significant if the confidence interval does not include 1 .0

In order to investigate the association between response on feeder, the behaviour and the day the 2x2 contingency tables were used and the difference between observed and expected frequency was tested by a χ² test.

RESULTS

Phase 1

During Phase 1 of the experiment the animals were trained to approach the dog feeder after the release of the sound. On the first day 31 .6 ±4.1 % of the animals were around the feeder. On day 4 the piglets had reached a rate of 50 ±3.6%. Subsequently, they never fell below that value and reached 71 ±3.3 % at the end of this experimental phase (day 8). This shows that the piglets learned quickly to recognize the dog feeder as a food source. Phase 2

From the whole video database of the five trials a total of 647 behavioral events were used for the analysis. Among the behaviours detected when the feeder was activated the most frequent were chase (189); fight (167) and attack with bite (162).

Response on the feeder sound.

The logistic regression showed that the type of behaviour had a significant effect (P<0.001 ) on the piglets response to the feeder sound (continuation or interruption of behaviour). The behaviours were included in the model as risk factors for the continuation of the behavioural event after the feeder sound exposure (Table 14). The results show the low risk of continuation of the specific aggressive behaviours such as head trust (OR=0.43; 95% C.I. 0.25-0.72), jump on other (0.56) or attack with bite (0.61 ) after the interruption. Ear biting had very low risk of continuation (0.55). Regarding the behaviours of elevated aggression level, the risk of continuation raised twice in a case of chase (OR=2.16; 95% C.I. 1 .13-2.2); fight had a seven times risk of being continued after the feeder sound release (OR=7.89; 95% C.I. 5.24-1 1 .89).

Table 14. Response to the feeder sound: odds ratios and 95 % confidence intervals (C.I.) for the analysed behaviours

The logistic regression model predicted the continuation of behavioural event with a probability rate of 0.28 when the feeder sound was released within the first second after it's start (Figure 9). The later the feeder sound was released after the start of the behavioural event the higher is the predicted probability that the action or fight continues. Chi-square analysis revealed a highly significant overall effect of the experimental day on the type of performed behaviours (χ²=102.2, DF=6, P<0.001 ). There was some indication that the intensity of aggression was greater during the first day of the experiment compared with the consecutive day (Figure 10). The frequency of the aggressive behaviours was higher during the first day, than during the second day (353 versus 263). During the first day a higher proportion of some specific aggressive behaviours occurred compared to the second day; such as fights (37% or 130 of 353 versus 14 % or 37 of 263) and chase (36% versus 24 %) (Figure 10). While on the second day piglets performed more attack with bite (36% versus 19%) and head trust (18% versus 6%). The number of ear biting did not differ for the two days.

The response of the piglets to the feeder sound (continued; interrupted) also differed significantly between the experimental days (χ²=129.6, DF=1 , P<0.001 ). On the second day the piglets interrupted 74.9% (197 of 263) of aggressive events, while on the first day of mixing just 33.7% (1 19 of 353) responded (Figure 10).

The response of the piglets to the feeder sound had a significant effect on the duration of behaviours (P<0.001 ). The reduction of the duration of each type of behaviour was noticed, when we compared interrupted aggressive behaviour by the feeder sound with continued aggressive behaviour after feeder activation (Table 15).

Table 15: Mean duration (s) (LSM ± SEM) of each type of behavioural event interrupted by the feeder sound versus continued.

Feeder latency

Table 16 shows the mean latency of approach to the feeder of the piglets which interrupted their behaviour after the feeder sound exposure was analysed. Even if the analysis showed a significant effect of the trial on the feeder latency (F=6.46, P>0.001 ), the reaction of the pig after the sound release did not delay more than 1 sec in all the experimental trials.

Table 16: The effect of the trial on the latency of response to the feeder sound (in seconds) of the piglets involved in aggressive interaction

The above results show that in the first phase of our experiments piglets were trained to associate the sound of the feeder with the release of the sweet feed potentially attractive for them through classical conditioning. Classical conditioning paradigm describes how neutral stimuli become conditioned through association, thus gaining the ability to elicit specific behaviours. In our case, we wanted to condition the piglets to rush immediately to the feeder when the feeder sound occurs. Our data of the performance in the training phase show that the piglets quickly recognize the dog feeder as a source of an attractive feed. At the third day the number of respondents doubled and remained on a high level during the consecutive days. These results indicate that three days of training before weaning could be enough to create an association in young piglets. About 30% of the piglets did not approach the feeder after the sound. Possibly they were afraid of novel stimuli or their ranking position in the group was so low that they avoided approaching the feeder area, when it was occupied by the penmates. Our results also indicate that in pigs sound and food enter fast into a functional association (3-4 days of training), leading to an approach of the animals to the feeder. A certain behaviour results when an effective stimulus is received or generated by the animal. When one behaviour occurs, an ongoing behaviour may be inhibited, if both behaviours cannot be performed at the same time. It is obvious that for the inhibition of an ongoing behaviour the new stimulus should be stronger than the current one. In the present study we evaluated whether the sweet feed stimulus is strong enough to inhibit aggressive or abnormal behaviour and can redirect the animal to the feeder. The results show that ear biting can be successfully interrupted (OR=0.55). Highly aggressive behaviours such as chase and fight were less likely to be interrupted (OR=7.89). The number of fights was drastically reduced on the second day (37 versus 130), when such behaviours as the short attacks with biting the opponent and head trusts occurred more frequently (96 versus 66 for attack with bite; 48 versus 23 for head trust). These behaviours were also found to be successfully interrupted by the feeder sound when applied quickly. The explanation could be that the majority of the hierarchical fights occurred already on the first day, while during the second day short aggressive events dominated, which probably just were tests of strength of dominant animals, which did not lead to any violent response of the receivers. It was also found that on the second day the feeder sound distracted the pigs from the majority of behaviours (73.6 %), while on the first day piglets were distracted only in 34.4 % of cases.

When a behaviour was successfully interrupted by the feeder sound, piglets redirected their attention immediately. The duration of interrupted behaviours was shorter compared to not interrupted behaviours. And also a general result appeared from the trials. The more time passed from the start of an aggression, the more the animals were involved in aggressive actions and the less was the probability to interrupt them by the sound signal.

In conclusion, the presented method bears is able to reduce the frequency and duration of aggressive or undesired actions among young piglets. When sufficiently trained the motivation for an attractive food bait can be in most of the cases (up to 74%) higher than to continue with a just started fight. Aggressive behaviours related to the establishment of a dominance hierarchy within a group are less likely to be interrupted. Aggressive and violent actions among young piglets in unstructured common pens of livestock production systems, caused by reasons different from hierarchy establishment, can be successfully reduced by a sound - food reward application.

Example 4. Reduction of the frequency and duration of aggressive behaviour - Resident-Intruder test

The same piglets that were trained in Example 3 (Phase 1 ) were used in a test called Resident-Intruder, where two piglets are confronted in a test arena for maximum 7 minutes, depending on their response towards each other. The test arena was formed by partitioning a portion of the home pen of a group of 12 trained piglets with a black board made of strong plastic. The resident piglet was first isolated in the arena built in its home pen. The intruder piglet was then collected from another pen and placed into the test arena already containing the resident piglet. If an attack occurred, the electronic feeder was activated in order to break the aggressive interaction. Pairs of piglets were randomly selected. 12 resident piglets were tested once a day with different partners for two days. In total 260 aggressive interactions were analysed in 3 rounds of experiment.

From 82 recorded videos, a total of 260 episodes of aggressive interactions were identified and manually labelled by an expert observer.

The observer wrote down which piglet started the aggressive interaction and as a

consequence who was the receiver, if they responded to the activation of the feeder or not, who was the first piglet to respond or both of them or none of the piglets responded by the fact that there was no fight or the fight did not stop.

In the Resident-Intruder test, the piglets played different roles by being aggressors or receivers during a certain confrontation. Regarding the aggressive interactions that could be effectively stopped, the receiver reacted in 65% and the aggressor reacted in 97% of the feeder activations (Figures 1 1 and 12).

In Figure 13, the type of reaction was divided in none, stopped and not stopped aggressive interaction and it was noticed that from the total number of encounters, 80% of the aggressive interactions could be stopped by the activation of the feeder.

When taking into account the kind of reaction of the piglets towards the training commands during the resident intruder test, it was observed that the aggressors reacted slightly more than the receivers (45% vs. 38%, respectively, Figure 14)

In the foregoing specification, the invention has been described with reference to specific examples of embodiments of the invention. It will, however, be evident that various modifications and changes may be made therein without departing from the broader spirit and scope of the invention as set forth in the appended claims.

The method and system according to the invention may instead for being used to monitor, modify or prevent undesired behavior of pigs, be used to monitor, modify or prevent undesired behaviour of other animals, such as poultry or cows. Undesired behavior amongst poulty may for instance comprise pecking order fights, feather picking, laying eggs in litter, water and feed spoilage and sexual dominance. Examples of undesired behaviour amongst cows may comprise aggression to other animals, humans, water and feed spoilage, sexual behaviour (mounting), lying on defecation areas, milking of other cows and navel licking. However, other modifications, variations and alternatives are also possible. The specifications, drawings and examples are, accordingly, to be regarded in an illustrative rather than in a restrictive sense.

In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word 'comprising' does not exclude the presence of other features or steps then those listed in a claim. Furthermore, the words 'a' and 'an' shall not be construed as limited to 'only one', but instead are used to mean 'at least one', and do not exclude a plurality. The mere fact that certain measures are recited in mutually different claims does not indicate that a combination of these measures cannot be used to advantage.

REFERENCES

Arey, D.S.; Franklin, M.F. (1995), Effects of Straw and Unfamiliarity on Fighting between Newly Mixed Growing Pigs. Applied Animal Behaviour Science, 45, 23-30)

Costa et al. (2009), Biosystems Engineering 104(1 ):1 18-124

D'Eath and Pickup (2002), Aggr Behav, 28, 401-415

Erhard et al., (1997), Applied Animal Behavioural Science. 54, 137-151

Gonyou et al, (1988), Journal of Animal Science. 66, 2856-2864

Hosmer and Lemeshow (1989), Applied Logistic Regression. New York: Wiley

Jensen (1980), Appl Anim Ethol, 6: 341 -350

Kleinbaum et al. (1982), Epidemiologic Research, Principles and Quantitative Methods. Van Nostrand Reinhold Company Inc., New York, NY, pp 314

Luescher, U.A.; Friendship, R.M.; McKeown, D.B. (1990) Canadian Journal of Animal Science. 70, 363-370

McGlone (1985) J. Anim. Sci., 61 , 559-565

O'Connell and Beattie (1999). Animal Welfare. 8, 269-279

Poletto, R. et al. (2010), Aggression in Replacement Grower and Finisher Gilts Fed a Short- Term High-Tryptophan Diet and the Effect of Long-Term Human-Animal Interaction. Applied Animal Behaviour Science. 122, 98-1 10)

Rushen and Pajor (1987) Aggr Behav, 13: 329-346

Tan, S.S.L.; Shackleton, D.M. (1990). Effects of Mixing Unfamiliar Individuals and of

Azaperone on the Social-Behaviour of Finishing Pigs. Applied Animal Behaviour Science. 26, 157-168

Taylor et al., 2010 Vet. J., doi: 10.1016/j.tvjl.2009.08.028

Turner et al. (2006) Applied Animal Behaviour Science, 2006, 96(3-4)

Zonderland, J.J. 2010. Talking tails - Quantifying the development of tail biting in pigs. 176

Claims

1 . An automated method of modifying or preventing undesired behaviour of a livestock animal or group of animals comprising the steps of: (i) automated monitoring of the behaviour of said animal or group of animals; (ii) generating a stimulus or trigger to prevent or stop the undesired behaviour.

2. The automated method according to claim 1 , further comprising the step of detecting or predicting the onset or early signs of undesired behaviour of said animal or group of animals, using the data generated in step (i).

3. The automated method according to claim 1 or 2, wherein said stimulus or trigger is generated when the early signs of undesired behaviour of said animal or group of animals have been detected or when a future undesired behaviour is predicted.

4. The automated method according to claims 1 to 3 further comprising the step of conditioning or training the animal to associate said trigger or stimulus with a reward, a punishment or warning.

5. The automated method according to claims 1 to 4 wherein said automated monitoring is a camera-based monitoring system.

6. The automated method according to any one of the preceding claims, wherein said animal or group of animals is a pig.

7. The automated method according to any one of claims 1 -6, wherein the data on animal behaviour is collected with means for monitoring the animal behaviour, said means comprising at least one suitable sensor chosen from a camera, a microphone, a motion sensor, a heat sensor, a heart rate sensor and a pressure sensor.

8. The automated method according to any one of the preceding claims, wherein real time data is compared to the collected data to determine the early signs of undesired behaviour.

9. The automated method according to any one of the preceding claims, wherein the trigger or stimulus comprises at least one of an audible, visual or sensory stimulus.

10. The automated method according to any one of the preceding claims, wherein collected data on initial behaviours, such as body positions of interacting animals or positions from distance without any contact, of the animal is labelled to relate this data to aggressive behaviour to enable identification of every labelled behaviour happening during a certain period of time.

1 1 . A method according to any one of the preceding claims, wherein the method further comprises collecting data for identifying individual livestock.

12. A system or device for controlling and preventing undesired behaviour of a livestock animal or group of animals, comprising means for monitoring animal behaviour, means for identifying or predicting the onset of undesired behaviour and means for generating a trigger to prevent or stop the undesired behaviour.

13. The system or device according to claim 12, wherein the means for monitoring the animal behaviour comprise at least one suitable sensor chosen from a camera, a microphone, a motion sensor, a heat sensor, a heart rate sensor and a pressure sensor configured to collect data on the animal behaviour.

14. The system or device according to any one of claims 12-13, wherein the monitoring system is coupled to a trigger/stimulus generating system for generating a trigger or stimulus to prevent or stop undesired behaviour.

15. The system or device according to any one of claims 12-14, further comprising a processing system capable of processing the data collected by the sensor(s) to identify, detect or predict undesired behaviour, preferably the onset or early signs of said undesired behaviour.

16. The system or device according to claim 15, wherein the processing system is configured to command said means for generating said trigger or stimulus.

17. The system or device according to claims 15 or 16, wherein said processing means comprises means for storage of data.

18. The system or device according to any one of claims 15-17, wherein the processing system is configured to operatively execute one or more algorithms or programs for analysing inputs received from the at least one sensor, said one or more algorithms being configured (A) (i) to identify the inputs the occurrence of undesired behaviour, particularly the onset or early signs of such undesired behaviour; or (ii) to predict future occurrence of such undesired behaviour; and once such undesired behaviour has been detected or predicted (B) to command the generation of a trigger or stimulus to stop such undesired behaviour.

19. The system or device according to any one of claims 12-18, wherein an individual animal is equipped with a tag, preferably a machine readable tag.