CN110263728A - Anomaly detection method based on improved pseudo- three-dimensional residual error neural network - Google Patents
Anomaly detection method based on improved pseudo- three-dimensional residual error neural network Download PDFInfo
- Publication number
- CN110263728A CN110263728A CN201910548528.5A CN201910548528A CN110263728A CN 110263728 A CN110263728 A CN 110263728A CN 201910548528 A CN201910548528 A CN 201910548528A CN 110263728 A CN110263728 A CN 110263728A
- Authority
- CN
- China
- Prior art keywords
- video
- abnormal
- neural network
- packet
- video clip
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of anomaly detection methods based on improved pseudo- three-dimensional residual error neural network, comprising the following steps: firstly, being multiple video clips by each Video segmentation in training set；Secondly, obtaining their feature after all video clips of a video are inputted improved pseudo- three-dimensional residual error neural network respectively；Then, the average value of the feature vector of all frames in each segment is taken, and the normalization of L2 norm and then is carried out to characteristic mean, to obtain the feature vector of the video clip；Finally, the feature vector of video clip to be input to one 3 layers of full Connection Neural Network, the abnormal score of video clip will be exported.The experimental results showed that mentioned method further improves the accuracy rate of unusual checking herein compared with current method, more fitting practical application.
Description
Technical field
The present invention relates to the anomaly detection methods under a kind of monitor video scene, more particularly to one kind to be based on more examples
The anomaly detection method of study and improved pseudo- three-dimensional residual error neural network, belongs to Video Analysis Technology field.
Background technique
Traditional video monitoring not only needs high mainly or by the abnormal behaviour artificially gone in monitoring scene
Human cost, and it is easy to produce visual fatigue, it results even in some abnormal behaviours and is not observed in time；Abnormal behaviour
Detection is intended to automatically detect that the exception in monitoring scene by video frequency signal processing and machine learning scheduling algorithm with analysis
Behavior, so that people be helped to take corresponding measure in time；Therefore, the unusual checking of monitoring scene has highly important
Research significance.
The research work of unusual checking early stage describes normal mode using the track characteristic of low layer, however due to difficulty
To obtain reliable track, these methods robustness and bad in the complicated or crowded scene for thering are many places to block；It considers
The deficiency of the space-time characteristic of track characteristic and low layer, histograms of oriented gradients (HOG), light stream histogram (HOF) and boundary histogram
Figure (MBH) is widely used, on this basis, Markov random field model (MRF), social force model (SFM), multiple dimensioned light
Stream histogram (MHOF), mixing dynamic texture (MDT) are proposed in succession；These methods are according to the training data of normal behaviour to just
Low probability mode detection is abnormal by normal behavior modeling, however, the feature of these engineers is difficult effectively to reflect behavior spy
Property, and calculate complexity.
With the success of rarefaction representation and dictionary learning method in some computer vision problems, researchers start to make
Learn the dictionary of normal behaviour with rarefaction representation, during the test, the mode with larger reconstructed error is considered different
Chang Hangwei；Recently, there is researcher using the self-encoding encoder study normal behaviour model based on deep learning, utilize reconstruct loss inspection
It surveys abnormal；Method is all based on such a it is assumed that i.e. any behavior for deviateing the normal behaviour mode learnt will all be considered as
It is abnormal；However, this hypothesis may be invalid, because normal behaviour and abnormal behaviour all have complicated variety, and it
Between boundary be sometimes it is fuzzy.
Only learn normal behaviour dictionary from the training data of normal behaviour, and is based on reconstructed error detection abnormal behaviour
Inappropriate, it is reasonable that normal behaviour and abnormal behaviour video data, which are all used, and should be as few as possible
It is carried out under mark information.
Summary of the invention
Technical problem to be solved by the present invention lies in the anomaly detection methods overcome under existing monitor video scene
Deficiency, provide a kind of based on improved pseudo- three-dimensional residual error neural network (Pseudo-3D Residual Neural
Network, abbreviation P3D-ResNet) anomaly detection method, P3D-ResNet is improved, and use it to study view
The feature of frequency.
Anomaly detection method of the present invention based on improved pseudo- three-dimensional residual error neural network, including following step
It is rapid:
Step 1 is based on multi-instance learning method, using only coarse grain scale designation (i.e. videl stage distinguishing label) training set,
Data prediction is carried out to each video in training set, is multiple video clips by each Video segmentation；
Step 2 improves P3D-ResNet network structure, and all video clips of a video are inputted respectively
Improved P3D-ResNet obtains the feature vector of each video frame in each video clip；
Step 3 calculates the average value of the feature vector of all frames in each video clip, and and then to characteristic mean
The normalization of L2 norm is carried out, to obtain the feature vector of the video clip；
The feature vector of step 3 video clip obtained is inputted one 3 layers of full Connection Neural Network by step 4
(fully connected neural network, abbreviation FC neural network), will export the abnormal score of video clip (i.e.
Belong to abnormal probability)；
Step 5 draws subject's operating characteristic curve (Receiver Operating Characteristic, abbreviation
ROC), corresponding area (Area Under ROC Curve, abbreviation AUC) under calculated curve carries out input video exception segment
Assessment.
Further, in step 1, each video in training set is considered as a packet, the video containing abnormal behaviour
It is marked as positive closure, not the labeled packet that is negative of the normal video of abnormal behaviour, adjusts the size and frame rate of the every frame of video, number
Data preprocess method are as follows: each video is divided into the video clip of the non-overlap containing same number of frames, each video clip conduct
Example in positive closure or negative packet.
Further, in step 2, the improvement of P3D-ResNet network structure is, in pseudo- Three dimensional convolution residual error nerve
On the basis of network frame, 3 × 3 × 3 average pondization operation joined in the fast part connection shortcuts, and each
It all joined batch normalization (batch normalization, abbreviation BN) operation after convolution operation.
Further, it for 3 layers of full Connection Neural Network of classifier used in step 4, is carried out using objective function
Training.
Further, in step 4, the design procedure of objective function are as follows:
1) in multi-instance learning algorithm, the training set with videl stage distinguishing label is represented by { (x1,y1),…,(xi,
yi),…,(xN,yN), wherein xiIt is i-th of packet in training set, yiFor the label of the packet, N is the total number wrapped in training set；
I-th of packet is represented by againX thereinikIndicate packet xiK-th of example, niFor positive closure xi
In example sum.Assuming that training is concentrated with NaA positive closure, then xi(i∈[1,Na]) it is positive closure, and yi=+1；Training set simultaneously
In have N-NaA negative packet, then xj(j∈[Na+ 1, N]) be negative packet, and yj=-1；
In the standard supervision classification problem using support vector machines (SVM), optimization aim are as follows:
It is 1. wherein risk item, is 2. regularization term, C is iotazation constant, and m is the sum of training example, yiIt is each
Exemplary label, φ (xi) image block or video clip feature vector, w is the weight parameter of model, and b is offset parameter.
2) because abnormal score is to belong to abnormal probability, abnormal video is higher than the abnormal score of normal video to be only
Reasonably, in order to improve the accuracy rate of detection, it is desirable to which abnormal video clip has higher exception point than normal video clip
Sequence loss can be used in number, it can make anomalous video segment obtain higher score than normal segment, however, in no video
In the case that fragment stage annotates (it is unknown for whether there is normal video clip in anomalous video), need to arrange using more examples
Sequence loss function:
Wherein xikAnd xjlRespectively indicate abnormal video clip and normal video clip, f (xik) and f (xjl) difference table
Show corresponding abnormal score, and value range is [0,1].
3) example of abnormal highest scoring is most likely to be really positive example (abnormal segment) in positive closure, bears abnormal in packet
The example of highest scoring is actually a normal example, but the negative example is likely to be falsely detected, so by claiming
Make the difficult example in abnormality detection.In order to solve problem above, we are not ranked up each example in packet, but strong
System is only ranked up two examples of highest scoring abnormal in positive packet and negative packet, in order to allow the positive example of abnormal highest scoring
It differs greatly on abnormality score with the negative example of abnormal highest scoring, using the sequence loss function of hinge-loss form:
4) only above to sort loss function not enough, it is also necessary in view of the sequential organization of video.Firstly, since segment
Sequence is continuously, so the abnormal score between two adjacent video clips should be relative smooth.In this regard, we are logical
Cross the abnormal score difference that addition timing smoothness constraint minimizes adjacent video clip.Secondly as occurring in video
The time range of abnormal behaviour is usually relatively small, so the score of example (video clip) should be sparse in positive closure；It is right
This, we are by addition sparsity constraints, so that the abnormal score of video clip has sparsity；In conjunction with above-mentioned smoothness constraint
And sparsity constraints, obtain complete objective function are as follows:
It is 1. wherein timing smoothness constraint term, is 2. sparse constraint, is 3. regularization term, w indicates Model Weight, λ1、λ2
And λ3Respectively indicate timing smoothness constraint term coefficient, sparse constraint term coefficient and regularization term coefficient, other variables and (2) formula phase
Together.In MIL sequence loss, error is the video clip backpropagation of the highest scoring from positive closure and negative packet.
Further, in step 5, the abnormal probability of each video clip is obtained using classifier, then draws ROC curve,
The performance of anomaly detection method can be assessed, method particularly includes:
1) descending sort is carried out to video clip according to abnormal probability, while using maximum abnormal probability as threshold value, then
It is classified as positive class more than or equal to the example of this threshold value, is otherwise classified as negative class；
2) false positive example rate and real example rate are calculated according to classification results, and is obtained respectively using them as abscissa, ordinate
One point of reference axis；
3) predicted value that classification thresholds are set gradually to other examples obtains a series of reference axis according still further to the above method
On point；
4) all coordinate points are connected into ROC curve, corresponding area is AUC under ROC curve.
The training data of the present invention for having the beneficial effect that the present invention and weak label being used only, is not only utilized normal row
For video data, also used the video data of abnormal behaviour, improved the accuracy rate of unusual checking, more fitting is real
Border application.
Detailed description of the invention
In order that the present invention can be more clearly and readily understood, right below according to specific embodiment and in conjunction with attached drawing
The present invention is described in further detail.
Fig. 1 is mentioned anomaly detection method flow chart by this paper；
Fig. 2 is the pseudo- three-dimensional residual error block structural diagram of 3 kinds in P3D-ResNet；
Fig. 3 is improved P3D-ResNet mount structure schematic diagram；
Fig. 4 is the ROC of the application and other two existing method and pedestal method based on three-dimensional nerve network frame
Curve；It wherein 1. indicates the method that this patent is proposed, 2. indicates the method based on 3D CNN, 3. indicate to be based on 3D ResNet-34
Method, 4. indicate the method based on two category support vector machines.
Specific embodiment
It is specific for based on more the object of the present invention is to provide the anomaly detection method under a kind of monitor video scene
Learn-by-example and the anomaly detection method of improved pseudo- three-dimensional residual error neural network are improved public with reinforcing monitoring capacity
Safety.
As shown in Figure 1, the implementation steps of the invention is as follows:
Firstly, each video in training set is divided into multiple video clips, and improved P3D-ResNet is inputted, obtained
Their feature；Then, the feature average value of all frames in each video clip is taken, L2 norm and then is carried out to characteristic mean
Normalization, to obtain the feature of the video；Finally, these features are input to one 3 layers of full Connection Neural Network, it will
Export the abnormal score of video clip.
3 kinds of pseudo- three-dimensional residual error block structures in residual error network constructed by the application are as shown in Fig. 2, due to pseudo- three-dimensional serial
Shortcut residual block (P3D-A), pseudo- three-dimensional parallel shortcut residual block (P3D-B), pseudo- three-dimensional serial parallel shortcut residual block (P3D-C) three
The performance that kind block structure mixes the three-dimensional residual error neural network (P3D-ResNet) of the puppet being arranged to make up in order is better than single
Three mutation network P3D-A ResNet, P3D-BResNet and P3D-C ResNet of block structure composition, therefore, the application is with 3
The pseudo- three-dimensional residual error block structure of kind sequentially replaces the 2 dimension residual error block structures of ResNet-50 in turn, obtains block structure diversified pseudo- three
Tie up convolution residual error neural network；On the basis of pseudo- three-dimensional residual error neural network framework, then in the fast connection part shortcuts
It joined 3 × 3 × 3 average pondization operation, and joined batch normalization operation after each convolution operation, Fig. 3 is provided
The schematic diagram of improved P3D-ResNet.
Specifically include the following contents:
1. the preparation of experimental data
Abnormal behaviour in previous training set is very single, and the training set having is that people performs and records on some position
Made of system, can not reflecting video monitoring real scene the case where.
Due to the limitation of former training set, the present invention uses the Large-Scale Training Data Set of a new videl stage distinguishing label
UCF-Anomaly-Detection-Dataset is come the method for assessing us.The training set is by not cropped monitor video group
At total duration is longer, shares 1900 videos, and wherein normal video and anomalous video are 950.The training set covers 13
The anomalous event of class real scene, including maltreat human or animal, arrest suspect, set on fire, hitting, traffic accident, burglary,
It explodes, have a fist fight, plundering, shooting incident, stealing, shoplifting and destroying his person property etc..
The training set is divided into two parts: training set includes 1610 videos (800 normal videos, 810 abnormal views
Frequently), test set includes 290 videos (150 normal videos, 140 anomalous videos)
2. experimental detail is arranged
All video frame pixels are uniformly adjusted to 240 × 320, frame speed is uniformly adjusted to 30 frame per second, to extract puppet three
Each video is divided into the video clip of several 16 frame lengths according to video length, and has 8 frame weights between continuous two segments by dimensional feature
It is folded.
Firstly, inputting video data is divided into multiple video clips；Secondly, these video clips are inputted improved puppet
Three-dimensional residual error neural network obtains feature；Then, the feature average value of all frames in each video clip is taken；And then to feature
Mean value carries out the normalization of L2 norm, obtains the character representation of inputting video data；Finally, obtained feature is inputted one 3 layers
Full Connection Neural Network, to realize that inputting video data exception segment detects.
Wherein the FC layer (full articulamentum) of first layer has 512 units, and the FC layer of the second layer and third layer has 32 respectively
With 1 unit, FC interlayer uses 50% Dropout regularization.The present embodiment is by ELU activation primitive and Swish activation primitive
It is respectively used to first layer and the full articulamentum of the last layer, and the use of initial learning rate is 0.001, β1=0.9, β2=0.999, ε=
1×10-8Adam optimizer；To obtain optimum performance, in the sequence loss of multi-instance learning, by timing smoothness constraint term system
Number, sparse constraint term coefficient and regularization term coefficient are set as λ1=λ2=8 × 10-5And λ3=0.01.
3. the selection of evaluation index
Common evaluation index in unusual checking is using corresponding under subject's operating characteristic curve (ROC) and curve
Area (AUC).
The horizontal axis of ROC curve is " false positive example rate " (False Positive Rate, abbreviation FPR), and the longitudinal axis is " real example
Rate " (True Positive Rate, abbreviation TPR):
Wherein TP, FN, FP, TN respectively indicate real example (true positive), false positive example (false positive),
True counter-example (true negative), the corresponding sample number of false counter-example (false negative) four kinds of situations.ROC curve more leans on
Close over, area under the curve AUC is bigger, then the accuracy rate detected is higher, and otherwise Detection accuracy is lower.
After the abnormal probability for obtaining all video clips using classifier, draws out ROC curve and calculate AUC, can comment
Estimate the performance of anomaly detection method.Method particularly includes: 1) descending sort is carried out to video clip according to abnormal probability, simultaneously
Using maximum abnormal probability as threshold value, then it is classified as positive class more than or equal to the example of this threshold value, is otherwise classified as negative class；2) according to
False positive example rate and real example rate are calculated according to classification results, and respectively using them as abscissa, ordinate, obtains one of reference axis
Point；3) classification thresholds are set gradually according to descending as remaining exemplary predicted value, can be obtained according to the method described above a series of
Point in reference axis；4) all coordinate points are connected into ROC curve, calculates corresponding area AUC under ROC curve.
The application and other two existing method and pedestal method based on three-dimensional nerve network are enterprising in UCF data set
It has gone unusual checking, and has drawn ROC curve and carry out Performance Evaluation, as shown in Figure 4.It is 1. proposed by the application in Fig. 4
The method based on improved pseudo- three-dimensional residual error neural network: be based on videl stage label data, pass through the more example methodologies of depth
Behavior pattern is practised, using improved pseudo- three-dimensional residual error neural network as feature extractor, to use 50% Dropout canonical
The three layers of full Connection Neural Network changed are as classifier, to realize the detection of abnormal behaviour；
2. being the method (3D CNN) based on the full Connection Neural Network of Three dimensional convolution: being based on videl stage label data, utilize 3
Dimension convolutional neural networks are characterized extractor, to use 3 layers of Dropout regularization of 50% full Connection Neural Network for classification
Device, to realize the detection of abnormal behaviour；
3. for the method for the three-dimensional residual error neural network (3D ResNet-34) based on 34 layers: being made using 3D ResNet-34
It is characterized extractor, with linear SVM (SVM) for classifier, to realize the detection of abnormal behaviour；
4. directly to carry out unusual checking using two category support vector machines (binary SVM) classifier, and should
Method is as pedestal method.
The application has carried out quantifying for unusual checking effect with the existing method based on three-dimensional nerve network frame
Comparison, as shown in table 1 below:
Table 1
| Method | Acc. | AUC |
| ①P3D-ResNet(ours) | 89.2 | 94.6 |
| ②C3D+FC | 76.6 | 86.5 |
| ③3D ResNet-34 | 68.9 | 72.7 |
| ④Binary classifier | 50.0 | 55.5 |
Can be seen that other three kinds of methods that are compared to from the experimental result of Fig. 4 and table 1, the mentioned method of the application it is different
Normal behavioral value effect is best.
The foregoing is merely preferred embodiments of the invention, are not intended as limitation of the invention further, all to utilize this
Various equivalence changes made by description of the invention and accompanying drawing content are within the scope of the present invention.
Claims (6)
1. the anomaly detection method based on improved pseudo- three-dimensional residual error neural network, which comprises the following steps:
Step 1 is based on multi-instance learning method, using the training set of only coarse grain scale designation (i.e. videl stage distinguishing label), to instruction
Practice each video concentrated and carry out data prediction, is multiple video clips by each Video segmentation；
Step 2 improves P3D-ResNet network structure, and all video clips of a video are inputted improvement respectively
P3D-ResNet, obtain the feature vector of each video frame in each video clip；
Step 3 calculates the average value of the feature vector of all frames in each video clip, and and then carries out to characteristic mean
L2 norm normalization, to obtain the feature vector of the video clip；
The feature vector of step 3 video clip obtained is inputted one 3 layers of full Connection Neural Network by step 4, will
Export the abnormal score of video clip；
Step 5, draws subject's operating characteristic curve, and corresponding area under calculated curve comments input video exception segment
Estimate.
2. the anomaly detection method according to claim 1 based on improved pseudo- three-dimensional residual error neural network, special
Sign is, in step 1, each video in training set is considered as a packet, the video containing abnormal behaviour is labeled to be positive
Packet, the labeled packet that is negative of the normal video of abnormal behaviour, does not adjust the size and frame rate of the every frame of video, data prediction side
Method are as follows: each video is divided between adjacent containing same number of frames to the video clip for having overlapping, each video clip is as just
Example in packet or negative packet.
3. the anomaly detection method according to claim 1 based on improved pseudo- three-dimensional residual error neural network, special
Sign is, in step 2, is to the improvement of P3D-ResNet network structure, in pseudo- Three dimensional convolution residual error neural network framework
On the basis of, it joined 3 × 3 × 3 average pondization operation in the fast part connection shortcuts, and after each convolution operation
It all joined batch normalization operation.
4. the anomaly detection method according to claim 1 based on improved pseudo- three-dimensional residual error neural network, special
Sign is, for 3 layers of full Connection Neural Network used in step 4, is trained with objective function.
5. the anomaly detection method according to claim 4 based on improved pseudo- three-dimensional residual error neural network, special
Sign is, in step 4, the design procedure of objective function are as follows:
1) in multi-instance learning algorithm, the training set with videl stage distinguishing label is expressed as { (x1,y1),…,(xi,
yi),…,(xi,yN), wherein xiIt is i-th of packet in training set, yiFor the label of the packet, N is the total number wrapped in training set,
I-th of packet is represented by againX thereinikIndicate packet xiK-th of example, niFor positive closure xi
In example sum；Assuming that training is concentrated with NaA positive closure, then xi(i ∈ [1, Na]) it is positive closure, and yi=+1；Training set simultaneously
In have N-NaA negative packet, then xj(j∈[Na+ 1, N]) be negative packet, and yj=-1；
In the standard supervision classification problem using support vector machines (SVM), optimization aim can be obtained are as follows:
It is 1. wherein risk item, is 2. regularization term, C is iotazation constant, and m is the sum of training example, and yi is each example
Label, φ (xi) image block or video clip feature vector, w is the weight parameter of model, and b is offset parameter；
2) using sequence loss, anomalous video segment is made to obtain higher abnormality score than normal segment, it is contemplated that not regarding
In the case that frequency fragment stage annotates, more examples sequence loss function is used:
Wherein xikAnd xjlRespectively indicate abnormal video clip and normal video clip, f (xik) and f (xjl) respectively indicate pair
The abnormal score answered, and value range is [0,1]；
3) each example in packet is not ranked up, but forced only to the two of highest scoring abnormal in positive packet and negative packet
A example is ranked up, in order to allow the positive example of abnormal highest scoring to differ very on abnormality score with the abnormal negative example of highest scoring
Greatly, using the sequence loss function of hinge-loss form:
4) sequential organization in view of video is also needed；Firstly, since fragment sequence is continuously, so two adjacent videos
Abnormal score between segment should be relative smooth, minimize adjacent video clip by adding timing smoothness constraint
Abnormal score difference；Secondly as the time range for being abnormal behavior in video is usually relatively small, so in positive closure
Exemplary score should be it is sparse, by add sparsity constraints so that the abnormal score of video clip have sparsity；Knot
It closes and states smoothness constraint and sparsity constraints, obtain complete objective function:
It is 1. wherein timing smoothness constraint term, is 2. sparse constraint, is 3. regularization term, w indicates Model Weight, λ1、λ2And λ3
Respectively indicate timing smoothness constraint term coefficient, sparse constraint term coefficient and regularization term coefficient.
6. the anomaly detection method according to claim 1 based on improved pseudo- three-dimensional residual error neural network, special
Sign is, in step 5, the abnormal probability of each video clip is obtained using classifier, draws ROC curve, method particularly includes:
1) descending sort is carried out to video clip according to abnormal probability, while using maximum abnormal probability as threshold value, be then greater than
Or it is classified as positive class equal to the example of this threshold value, otherwise it is classified as negative class；
2) false positive example rate and real example rate are calculated according to classification results, and obtains coordinate using them as abscissa, ordinate respectively
One point of axis；
3) predicted value that classification thresholds are set gradually to other examples obtains in a series of reference axis according still further to the above method
Point；
4) all coordinate points are connected into ROC curve, calculates corresponding area AUC under ROC curve.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910548528.5A CN110263728B (en) | 2019-06-24 | 2019-06-24 | Abnormal behavior detection method based on improved pseudo-three-dimensional residual error neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910548528.5A CN110263728B (en) | 2019-06-24 | 2019-06-24 | Abnormal behavior detection method based on improved pseudo-three-dimensional residual error neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110263728A true CN110263728A (en) | 2019-09-20 |
| CN110263728B CN110263728B (en) | 2022-08-19 |
Family
ID=67920872
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910548528.5A Active CN110263728B (en) | 2019-06-24 | 2019-06-24 | Abnormal behavior detection method based on improved pseudo-three-dimensional residual error neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110263728B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111159961A (en) * | 2019-12-13 | 2020-05-15 | 杭州广立微电子有限公司 | Abnormity detection method and system based on curve data |
| CN111160117A (en) * | 2019-12-11 | 2020-05-15 | 青岛联合创智科技有限公司 | Abnormal behavior detection method based on multi-example learning modeling |
| CN111178344A (en) * | 2020-04-15 | 2020-05-19 | 中国人民解放军国防科技大学 | Multi-scale time sequence behavior identification method |
| CN111291699A (en) * | 2020-02-19 | 2020-06-16 | 山东大学 | Substation personnel behavior identification method based on monitoring video time sequence action positioning and abnormity detection |
| CN111739492A (en) * | 2020-06-18 | 2020-10-02 | 南京邮电大学 | Music melody generation method based on pitch contour curve |
| CN111814548A (en) * | 2020-06-03 | 2020-10-23 | 中铁第四勘察设计院集团有限公司 | Abnormal behavior detection method and device |
| CN111814653A (en) * | 2020-07-02 | 2020-10-23 | 苏州交驰人工智能研究院有限公司 | Method, device, equipment and storage medium for detecting abnormal behaviors in video |
| CN112257545A (en) * | 2020-10-19 | 2021-01-22 | 安徽领云物联科技有限公司 | Violation real-time monitoring and analyzing method and device and storage medium |
| CN113312968A (en) * | 2021-04-23 | 2021-08-27 | 上海海事大学 | Real anomaly detection method in surveillance video |
| CN113627386A (en) * | 2021-08-30 | 2021-11-09 | 山东新一代信息产业技术研究院有限公司 | Visual video abnormity detection method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108734095A (en) * | 2018-04-10 | 2018-11-02 | 南京航空航天大学 | A kind of motion detection method based on 3D convolutional neural networks |
| CN109117774A (en) * | 2018-08-01 | 2019-01-01 | 广东工业大学 | A kind of multi-angle video method for detecting abnormality based on sparse coding |
| CN109635790A (en) * | 2019-01-28 | 2019-04-16 | 杭州电子科技大学 | A kind of pedestrian's abnormal behaviour recognition methods based on 3D convolution |
| CN109670446A (en) * | 2018-12-20 | 2019-04-23 | 泉州装备制造研究所 | Anomaly detection method based on linear dynamic system and depth network |
-
2019
- 2019-06-24 CN CN201910548528.5A patent/CN110263728B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108734095A (en) * | 2018-04-10 | 2018-11-02 | 南京航空航天大学 | A kind of motion detection method based on 3D convolutional neural networks |
| CN109117774A (en) * | 2018-08-01 | 2019-01-01 | 广东工业大学 | A kind of multi-angle video method for detecting abnormality based on sparse coding |
| CN109670446A (en) * | 2018-12-20 | 2019-04-23 | 泉州装备制造研究所 | Anomaly detection method based on linear dynamic system and depth network |
| CN109635790A (en) * | 2019-01-28 | 2019-04-16 | 杭州电子科技大学 | A kind of pedestrian's abnormal behaviour recognition methods based on 3D convolution |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160117A (en) * | 2019-12-11 | 2020-05-15 | 青岛联合创智科技有限公司 | Abnormal behavior detection method based on multi-example learning modeling |
| CN111159961A (en) * | 2019-12-13 | 2020-05-15 | 杭州广立微电子有限公司 | Abnormity detection method and system based on curve data |
| CN111159961B (en) * | 2019-12-13 | 2024-02-02 | 杭州广立微电子股份有限公司 | Anomaly detection method and system based on curve data |
| CN111291699A (en) * | 2020-02-19 | 2020-06-16 | 山东大学 | Substation personnel behavior identification method based on monitoring video time sequence action positioning and abnormity detection |
| CN111178344A (en) * | 2020-04-15 | 2020-05-19 | 中国人民解放军国防科技大学 | Multi-scale time sequence behavior identification method |
| CN111814548A (en) * | 2020-06-03 | 2020-10-23 | 中铁第四勘察设计院集团有限公司 | Abnormal behavior detection method and device |
| CN111739492B (en) * | 2020-06-18 | 2023-07-11 | 南京邮电大学 | Music melody generation method based on pitch contour curve |
| CN111739492A (en) * | 2020-06-18 | 2020-10-02 | 南京邮电大学 | Music melody generation method based on pitch contour curve |
| CN111814653A (en) * | 2020-07-02 | 2020-10-23 | 苏州交驰人工智能研究院有限公司 | Method, device, equipment and storage medium for detecting abnormal behaviors in video |
| CN111814653B (en) * | 2020-07-02 | 2024-04-05 | 苏州交驰人工智能研究院有限公司 | Method, device, equipment and storage medium for detecting abnormal behavior in video |
| CN112257545A (en) * | 2020-10-19 | 2021-01-22 | 安徽领云物联科技有限公司 | Violation real-time monitoring and analyzing method and device and storage medium |
| CN113312968A (en) * | 2021-04-23 | 2021-08-27 | 上海海事大学 | Real anomaly detection method in surveillance video |
| CN113312968B (en) * | 2021-04-23 | 2024-03-12 | 上海海事大学 | Real abnormality detection method in monitoring video |
| CN113627386A (en) * | 2021-08-30 | 2021-11-09 | 山东新一代信息产业技术研究院有限公司 | Visual video abnormity detection method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110263728B (en) | 2022-08-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110263728A (en) | Anomaly detection method based on improved pseudo- three-dimensional residual error neural network | |
| CN106874894B (en) | Human body target detection method based on regional full convolution neural network | |
| WO2017190574A1 (en) | Fast pedestrian detection method based on aggregation channel features | |
| CN109902573B (en) | Multi-camera non-labeling pedestrian re-identification method for video monitoring under mine | |
| WO2018107760A1 (en) | Collaborative deep network model method for pedestrian detection | |
| CN109492581A (en) | A kind of human motion recognition method based on TP-STG frame | |
| CN107230267B (en) | Intelligence In Baogang Kindergarten based on face recognition algorithms is registered method | |
| Wang et al. | Detecting semantic parts on partially occluded objects | |
| CN108416250A (en) | Demographic method and device | |
| CN107133569A (en) | The many granularity mask methods of monitor video based on extensive Multi-label learning | |
| Feng et al. | Deep representation for abnormal event detection in crowded scenes | |
| CN104063719A (en) | Method and device for pedestrian detection based on depth convolutional network | |
| CN103390164A (en) | Object detection method based on depth image and implementing device thereof | |
| CN107220603A (en) | Vehicle checking method and device based on deep learning | |
| CN113920107A (en) | Insulator damage detection method based on improved yolov5 algorithm | |
| CN109190475A (en) | A kind of recognition of face network and pedestrian identify network cooperating training method again | |
| CN108052865A (en) | A kind of flame detecting method based on convolutional neural networks and support vector machines | |
| Nojiri et al. | Discussion on machine learning and deep learning based makeup considered eye status recognition for driver drowsiness | |
| Rentao et al. | Indoor smoking behavior detection based on yolov3-tiny | |
| CN111079694A (en) | Counter assistant job function monitoring device and method | |
| CN115512387A (en) | Construction site safety helmet wearing detection method based on improved YOLOV5 model | |
| CN106991400A (en) | A kind of fire hazard smoke detecting method and device | |
| Zambanini et al. | Detecting falls at homes using a network of low-resolution cameras | |
| CN114782979A (en) | Training method and device for pedestrian re-recognition model, storage medium and terminal | |
| Zhao et al. | Hybrid generative/discriminative scene classification strategy based on latent Dirichlet allocation for high spatial resolution remote sensing imagery |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |