CN204946028U - Object detecting device - Google Patents

Abstract

The utility model discloses a kind of object detecting device, described device comprises: characteristic pattern checkout equipment, filter apparatus and classification and Detection equipment.The utility model is detected by deformable part model and multiple SVM classifier realize target, can not only significantly promote detection efficiency, accuracy of detection, accuracy of detection and performance can meet actual requirement, and are expected to the application reaching heavy industrialization in a short time.The aspects such as vehicle flowrate, people flow rate statistical, vehicle security drive, rule-breaking vehicle detection can be applied in.

Description

Object detecting device

Technical field

The utility model relates to target detection technique field, particularly relates to a kind of object detecting device.

Background technology

At present, target detection is academicly having many models, can deformable member model (DeformablePartModel, DPM) be wherein object detection model in the most popular recently figure, receiving an acclaim owing to detecting target, is generally acknowledged best algorithm of target detection.

But, use at present the target detection technique of DPM be applied to vehicle, pedestrian detection time, its Detection results still needs to promote.Existing target detection technique detection efficiency, accuracy of detection not only fail to reach actual requirement, and are difficult to heavy industrialization application.

Summary of the invention

For the technical matters solving existing target detection technique detection efficiency, accuracy of detection fails to reach actual requirement, the utility model embodiment provides a kind of object detecting device.

For achieving the above object, the technical scheme of the utility model embodiment is achieved in that

A kind of object detecting device, described device comprises:

Characteristic pattern checkout equipment, obtains the characteristic pattern of input picture and twice resolution characteristics figure and is transferred to filter apparatus;

Filter apparatus, comprises a root wave filter and multiple parts wave filter; Described wave filter receives characteristic pattern that described characteristic pattern checkout equipment transmits and acts on the global characteristics described characteristic pattern capturing target object, and described global characteristics is transferred to classification and Detection equipment; Parts wave filter described in each receives the characteristic pattern of described two resolutions that described characteristic pattern checkout equipment transmits and the characteristic pattern acting on described two resolutions captures the local feature of target object, and described local feature is transferred to described classification and Detection equipment;

Classification and Detection equipment, comprise the multiple SVM classifier based on EM Algorithm for Training, for receiving global characteristics and the local feature of the target object that described filter apparatus transmits, and according to the global characteristics of described target object and local feature, identify the target object in described input picture, output detections result.

Wherein, described wave filter and described multiple parts wave filter form hub-and-spoke configuration.

Wherein, described parts wave filter position-movable, described local feature comprises the morphological change characteristics of target object.

Wherein, described classification and Detection equipment adopts based on EM Algorithm for Training mixing SVM model, by the mutual compensating action between multiple Linear SVM model, the target object that need detect is divided into multiple subclass according to the difference of visual angle or shape, for each subclass builds corresponding SVM model, obtain multiple SVM classifier, to detect the target object under various different visual angles.

Wherein, described filter apparatus is the filter apparatus adopting weighting deformable part model.

The object detecting device that the utility model provides is detected by deformable part model and multiple SVM classifier realize target, can not only significantly promote detection efficiency, accuracy of detection, accuracy of detection and performance can meet actual requirement, and are expected to the application reaching heavy industrialization in a short time.The aspects such as vehicle flowrate, people flow rate statistical, vehicle security drive, rule-breaking vehicle detection can be applied in.

Accompanying drawing explanation

In accompanying drawing (it is not necessarily drawn in proportion), similar Reference numeral can describe similar parts in different views.The similar reference numerals with different letter suffix can represent the different examples of similar parts.Accompanying drawing generally shows each embodiment discussed herein by way of example and not limitation.

Fig. 1 is the composition structural representation of the utility model object detecting device;

Fig. 2 is the utility model training process schematic diagram to mixing SVM under EM algorithm frame;

Fig. 3 is that the multiple SVM of the utility model divides schematic diagram to the son of primal problem;

Fig. 4 is the detection mean accuracy schematic diagram of the utility model mixing SVM model when testing on PASCALVOC2007 data set in 20 classes.

Embodiment

The utility model embodiment provides a kind of object detecting device, and as shown in Figure 1, described device comprises:

Characteristic pattern checkout equipment 11, obtains the characteristic pattern of input picture and twice resolution characteristics figure and passes to filter apparatus;

Filter apparatus 12, comprise a root wave filter and multiple parts wave filter, described wave filter acts on described characteristic pattern, capture the global characteristics of target object and described global characteristics is passed to classification and Detection equipment, parts wave filter described in each acts on the characteristic pattern of described two resolutions, captures the local feature of target object and described local feature is passed to described classification and Detection equipment;

Classification and Detection equipment 13, comprises the multiple SVM classifier based on EM Algorithm for Training, for according to the described global characteristics of target object and local feature, identifies the target object in described input picture, and output detections result.

The object detecting device of the utility model embodiment can be applicable in the statistical item of vehicle flowrate, and its accuracy of detection and performance reach actual requirement, and is expected to the application reaching heavy industrialization in a short time.It also can be applied in the aspect such as safe driving, rule-breaking vehicle detection of people flow rate statistical, automobile simultaneously.

The specific implementation of the present embodiment to the utility model object detecting device is described in detail.

The object detecting device of the utility model embodiment mainly based on technology such as the deformable part model of weighting and EM training mixing SVM, the detection of the similar object such as pedestrian and vehicle achieves good effect.

Wherein, in object detecting device, filter apparatus adopts deformable part model, this deformable part model is the Star Model be made up of a root wave filter and some parts wave filters, this deformable part model has considered the Global Information of target and the presentation information of each several part and spatial relationship thereof, can extract the information more enriched than the method based on entirety.

The principal feature of this deformable part model has following 2 points:

(1) the feature interpretation ability that combines of overall situation and partial situation, root wave filter is a kind of overall template to target, the global characteristics of target can be caught, and parts wave filter is a kind of local template, and act on the characteristic pattern of two resolutions, effectively can describe the local feature of target;

(2) for the descriptive power of target morphology change, because root wave filter is rigidity, do not possess the ability of process deformation, and the position of parts wave filter is transportable in deformable part model, and this characteristic can the metamorphosis of the target such as human body effectively.

Therefore, the hub-and-spoke configuration adopting root wave filter and parts wave filter to form can promote detection perform effectively.

The deformable component module adopted in the utility model object detecting device is weighting deformable component module, and this weighting deformable model has following 2 detection perform improved in order to promote deformable part model on the basis of existing deformable part model:

1) weighting block model.The power of comprehensive analysis all parts classifier performance, can strengthen the classification performance of marking area by weighting block, strengthen the utilization of some parts in Detection results.

Here, weighting block model is the deformable part model of weighting, the weight in addition different to different parts, being equivalent to be multiplied by a weight to the detection score that each parts of deformable part model are corresponding makes more effective parts be subject to more attention, the effect of maximized performance all parts, thus obtain better Detection results.

2) because single Linear SVM model can not solve the classification of two classification problems effectively, and the computation complexity of nonlinear SVM model is very high, is difficult to reach the object detected in real time.The utility model embodiment adopts by EM Algorithm for Training mixing SVM model.

EM algorithm, i.e. EM algorithm (ExpectationMaximizationAlgorithm, translate expectation-maximization algorithm again), a kind of iterative algorithm, for maximal possibility estimation or the maximum a posteriori estimate of the probability parameter model containing hidden variable (hiddenvariable).In statistical computation, EM algorithm is the algorithm finding parameter maximal possibility estimation or MAP estimation in probability (probabilistic) model, and wherein probability model depends on the hidden variable (LatentVariable) that cannot observe.Greatest hope is through being commonly used in data clusters (DataClustering) field of machine learning and computer vision.

EM algorithm to hocket calculating through two steps: the first step is calculation expectation (E), utilizes the existing estimated value to hidden variable, calculates its maximum likelihood estimator; Second step maximizes (M), maximizes the value that the maximum likelihood value of trying to achieve in E step carrys out calculating parameter.The estimates of parameters that M step finds is used to next E and walks in calculating, and this process constantly hockets, and iteration uses EM step, until convergence.

In the utility model embodiment, classification and Detection equipment adopts based on the multiple Linear SVM model of EM Algorithm for Training, by the mutual compensating action between multiple Linear SVM model, till being continued until convergence.And the key issue of this model is the mutual and distribution method between each submodel, utilize this mixing of EM Algorithm for Training SVM model, the output of SVM is converted into probability, then this mixing of framework joint training SVM model of EM is utilized, can make so mutually to compensate between the detection perform of each subclass model, demonstrate its convergence and validity by experiment.

In the utility model embodiment, classification and Detection equipment is by the mutual compensating action between multiple Linear SVM model, the target that need detect is divided into some subclasses according to the difference of visual angle or shape, make the similarity between each subclass very high, then be that each subclass builds corresponding detection model, difference can be overcome in other class of target class on the impact of detection perform by the combination of multiple SVM model, thus accurately detect the target under various different visual angles, reach the target promoting Detection results.

Fig. 2 and Fig. 3 show respectively classification and Detection equipment by multiple SVM of primal problem divided and under EM algorithm frame to the training process of mixing SVM and classification results.

As shown in Figure 2, the training process mixing SVM in classification and Detection equipment comprises the steps, wherein input comprises positive sample P, negative sample N and initialization model ω, β, exports and comprises new detection mathematical model β _new, ω _new:

Step 201: first all samples are divided into M class before training, every class sorter ω, β do follow-up training.For all positive sample set (x _i, y _i) perform following iterative step:

Step 202: calculate the sample of input by SVM classifier, the model that the result of output describes according to following formula (1) carries out probability conversion.

$pk\left(x_i\right)=\frac{{\mathrm{\π}}_k{\mathrm{\φ}}_k\left(x_i\right)}{{\Σ}_{k=1}^K{\mathrm{\π}}_k{\mathrm{\φ}}_k\left(x_i\right)}---\left(1\right)$

Step 203: perform M operation;

For each sample, according to pk (x _i) value sort, maximum pk (x _i) the sample set at positive sample place will deliver to step 4 and carry out iterative computation.

Step 204: for the sample set calculated by EM method, calculates weighted value α by following formula (2), (3);

${\mathrm{\ρ}}_{kp}=\frac{\left|\right|w_{kp}|{|}_2}{{\mathrm{\Σ}}_{p=1}^P\left|\right|w_{kp}|{|}_2}---\left(2\right)$

${\mathrm{\α}}_{kp}=\frac{e^{\mathrm{\λ}.{\mathrm{\ρ}}_{kp}}}{{\mathrm{\Σ}}_{p=1}^P{e}^{e^{\mathrm{\λ}.{\mathrm{\ρ}}_{kp}}}}\·P.---\left(3\right)$

Step 205: obtain the parameter ω, the β that sample set are trained weighting block SVM in EM method, obtains new detection mathematical model β _new, ω _new.

The object detecting device of the utility model embodiment is applied to the detection of upright pedestrian, its specific implementation process can comprise:

Step a1, input picture is to characteristic pattern checkout equipment;

Step a2, characteristic pattern checkout equipment obtains the characteristic pattern on the characteristic pattern of input picture and two resolutions, and the characteristic pattern of described characteristic pattern and two resolutions is transferred to filter apparatus;

Step a3, the root wave filter of filter apparatus receives described characteristic pattern and acts on described characteristic pattern, obtains the response image of root wave filter, and described response image is transferred to classification and Detection equipment; The all parts wave filter of filter apparatus receives the characteristic pattern of described two resolutions and the characteristic pattern acted on described two resolutions, obtain the response image of parts wave filter, and the response obtained further after range conversion, the response after response image and range conversion is all transferred to classification and Detection equipment;

Step a4, classification and Detection equipment accepts filter response image that root filter transfer in equipment comes and the response image that all parts filter transfer is come, and the response image that the response image of being come by described filter transfer and all parts filter transfer are come merges, obtain the overall response of input picture, then to be identified the personage in image by multiple SVM classifier based on described overall response;

Step a5, classification and Detection equipment is by testing result and export.

Here, in testing result, iris out each personage detected in the picture in the mode of picture frame.

The object detecting device of the utility model embodiment has been successfully applied in wagon flow quantitative statistics, and its accuracy of detection and performance reach actual requirement, and is expected to the application reaching heavy industrialization in a short time.Also can be applied to the aspect such as safe driving, rule-breaking vehicle detection of people flow rate statistical, automobile simultaneously.

Mixing SVM model after improvement is tested on PASCALVOC2007 data set, the performance of the object detecting device of the utility model embodiment improves 2.1%mAP on the whole under identical testing conditions, the mean accuracy form as shown in Figure 4 of the detection specifically in 20 classes.

The above, be only preferred embodiment of the present utility model, is not intended to limit protection domain of the present utility model.

Claims (5)

1. an object detecting device, is characterized in that, described device comprises:

Characteristic pattern checkout equipment, obtains the characteristic pattern of input picture and twice resolution characteristics figure and is transferred to filter apparatus;

Filter apparatus, comprises a root wave filter and multiple parts wave filter; Described wave filter receives characteristic pattern that described characteristic pattern checkout equipment transmits and acts on the global characteristics described characteristic pattern capturing target object, and described global characteristics is transferred to classification and Detection equipment; Parts wave filter described in each receives the characteristic pattern of described two resolutions that described characteristic pattern checkout equipment transmits and the characteristic pattern acting on described two resolutions captures the local feature of target object, and described local feature is transferred to described classification and Detection equipment;

Classification and Detection equipment, comprise the multiple SVM classifier based on EM Algorithm for Training, for receiving global characteristics and the local feature of the target object that described filter apparatus transmits, and according to the global characteristics of described target object and local feature, identify the target object in described input picture, output detections result.

2. device according to claim 1, is characterized in that, described wave filter and described multiple parts wave filter form hub-and-spoke configuration.

3. device according to claim 1 and 2, is characterized in that,

Described parts wave filter position-movable, described local feature comprises the morphological change characteristics of target object.

4. device according to claim 1, is characterized in that:

Described classification and Detection equipment adopts based on EM Algorithm for Training mixing SVM model, by the mutual compensating action between multiple Linear SVM model, the target object that need detect is divided into multiple subclass according to the difference of visual angle or shape, for each subclass builds corresponding SVM model, obtain multiple SVM classifier, to detect the target object under various different visual angles.

5. device according to claim 1, is characterized in that: described filter apparatus is the filter apparatus adopting weighting deformable part model.