CN105631463A - Time-space movement profile feature-based pedestrian detection method - Google Patents

Abstract

The invention belongs to the video image processing technical field and provides a time-space movement profile feature-based pedestrian detection method. The method includes the following steps that: S1, a 3D Haar filter is utilized to extract the 3D Haar feature vectors of a time-space body in a training sample set; S2, the 3D Haar feature vectors are trained based on a Gentle Adaboost cascade algorithm, so that a pedestrian cascade classifier can be obtained; S3, the pedestrian cascade classifier is utilized to perform pedestrian detection on input images acquired by a self monitoring region; and S4, pedestrians detected in the step S3 is tracked and counted based on an EKM algorithm. According to the method of the invention, the 3D Haar feature vectors in the training sample set are calculated, and the optimized pedestrian cascade classifier is utilized to detect pedestrians in the monitoring region, and therefore, the efficiency and accuracy of counting of pedestrians in a public region can be effectively improved.

Description

A kind of pedestrian detection method based on space-time body motion outline feature

Technical field

The invention belongs to Video Image processing technology field, particularly to a kind of pedestrian detection method based on space-time body motion outline feature, for the pedestrian's quantity in public territory is carried out accurate count.

Background technology

Along with the development of computer technology and image processing techniques, the intelligent monitor system based on video is widely used. In ensureing social public security and traffic safety, protection people life property safety, industrial control field guarantee safe production and in Product checking and about commercial field in all play huge effect. At present, the application of intelligent video monitoring system is mainly in security field and non-security prevention and control field. The monitoring of public place population surveillance, traffic safety, commercial production security monitoring etc. broadly fall into the application in security field. Secure world has: the detection of commercial field, industrial products, public transportation system etc.

The statistical information of number flow has important effect for a lot of industries, they can utilize flow of the people information assists to manage, reasonably configuration human and material resources are thus efficiently utilizing limited resource, or reasonably control crowd density according to the people information of statistics and prevent the overcrowding of crowd from security incident occurring. Such as large-scale stadium, exhibition center, heavy construction, number information can help reasonable (such as: seat, public health infrastructure etc.) that whether service facility that administration section's assessment provides is enough, whether convenient, whether build, thus configuration resource improves the utilization rate of Architectural Equipment in good time. People information also can be supplied to building design units and provide reference information for building appropriate design.

But, in actual use, based on the prior art of one-dimensional description of pedestrian contour in recognition efficiency and accuracy of detection all not satisfactory. Therefore, it is necessary to propose to improve to pedestrian detection method of the prior art.

Summary of the invention

It is an object of the invention to openly a kind of pedestrian detection method based on time-space domain motion contour feature, in order to improve in public territory, pedestrian is carried out demographics efficiency and accuracy.

For achieving the above object, the invention provides a kind of pedestrian detection method based on space-time body motion outline feature, comprise the following steps:

S1, the three-dimensional Haar wave filter of use extract the three-dimensional Haar characteristic vector of the space-time body that training sample is concentrated;

S2, based on GentleAdaboost Cascade algorithms, described three-dimensional Haar characteristic vector is trained, obtains pedestrian's cascade classifier;

S3, utilize pedestrian's cascade classifier to from monitoring region obtain input picture carry out pedestrian detection;

S4, based on EKM algorithm the pedestrian detected in step S3 it is tracked and counts.

As a further improvement on the present invention, described training sample set includes positive sample set and negative sample collection, described positive sample set is made up of some positive samples comprising pedestrian area image, and described negative sample collection is made up of some negative samples not comprising and/or not exclusively comprising pedestrian area image.

As a further improvement on the present invention, described step S1 particularly as follows:

S11. the 5 adjacent positive samples of frame or negative sample are defined as space-time body go forward side by side row space and time gradient conversion, so that the profile of space-time body is converted into range conversion space-time body;

S12. conversion space-time body of adjusting the distance is integrated computing and obtains integration body, and calculates the pixel value sum in integration body;

S13. utilize three-dimensional Haar wave filter to carry out convolution with integration body, obtain the three-dimensional Haar characteristic vector of space-time body.

As a further improvement on the present invention, the positive sample in described step S11 or negative sample to be pixel size be 30 �� 30 256 rank gray level images.

As a further improvement on the present invention, the integral operation formula in described step S12 is: $\mathrm{IV}(x,y,t)=\underset{x^{\&prime;}<x,y^{\&prime;}<y,t^{\&prime;}<t}{\mathrm{\&Sigma;}}D(x^{\&prime;},y^{\&prime;},t^{\&prime;});$

Wherein, (x, y, t) for integration body, D (x ', y ', t ') is range conversion space-time body to IV.

As a further improvement on the present invention, the region of described integration body is the space structure being included in range conversion space-time body, and described integration body is at (x, y, t) the pixel value sum at place is that in space-time body, all coordinate figures are respectively less than (x, y, t) the pixel value sum at place.

As a further improvement on the present invention, three-dimensional Haar wave filter in described step S13 is by being used for extracting three static Haar wave filter S1, S2, S3 of pedestrian's static exercise feature, and four dynamic Haar filter Ds 1, D2, D3, D4 for extracting pedestrian's behavioral characteristics collectively constitute.

As a further improvement on the present invention, described " from the input picture that monitoring region is acquired " in step S3 is the video streaming image being obtained monitoring region by video camera, and described monitoring region is positioned at the underface of video camera.

As a further improvement on the present invention, described " EKM algorithm " in step S4 is particularly as follows: according to the coordinate of former frame target location in continuous print input picture, utilize Kalman filter to predict the coordinate points that in this frame input picture, target is possible, then utilize meanshift algorithm to be iterated computing with the coordinate points estimated for starting point.

Compared with prior art, the invention has the beneficial effects as follows: by calculating the three-dimensional Haar characteristic vector that training sample is concentrated, and by the pedestrian's cascade classifier that utilizes optimized, the pedestrian in monitoring region is detected, it is effectively improved in public territory, pedestrian is carried out demographics efficiency and accuracy.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of a kind of pedestrian detection method based on space-time body motion outline feature of the present invention;

Fig. 2 is the schematic diagram of seven kinds of three-dimensional Haar wave filter of the present invention;

Fig. 3 is the schematic diagram of the integration body in range conversion space-time body;

Fig. 4 obtains the video streaming image schematic diagram as input picture from monitoring region in step S3;

Fig. 5 utilizes pedestrian's cascade classifier that input picture is carried out the schematic diagram of pedestrian detection in step S3.

Detailed description of the invention

Below in conjunction with each embodiment shown in the drawings, the present invention is described in detail; but it should what illustrate is; these embodiments are not limitation of the present invention; those of ordinary skill in the art, according to the equivalent transformation in these embodiment institute work energy, method or structure or replacement, belong within protection scope of the present invention.

The detailed description of the invention of the ginseng a kind of pedestrian detection method based on time-space domain motion contour feature of the present invention shown in Fig. 1. Owing to the change of pedestrian's head when walking Yu shoulder is less, based on the consideration being prone to context of detection, training sample set can be defined as: only comprise the positive sample set of wardrobe portion and/or shoulder, do not comprise the negative sample atlas of pedestrian head and/or shoulder.

In the present embodiment, step S1 is first carried out, uses three-dimensional Haar wave filter to extract the three-dimensional Haar characteristic vector of the space-time body that training sample is concentrated.

Concrete, this step S1 specifically includes following sub-step.

Sub-step S11. the 5 adjacent positive samples of frame or negative sample are defined as space-time body go forward side by side row space and time gradient conversion, so that the profile of space-time body is converted into range conversion space-time body.

In the present embodiment, this training sample set includes positive sample set and negative sample collection; Wherein, the 256 rank gray level images that positive/negative sample is 30 �� 30 pixels in this positive/negative sample set, described positive sample is the image comprising pedestrian area, and described negative sample is the image not comprising or not exclusively comprising pedestrian area. Further, the so-called image not comprising pedestrian area, refer to the image not comprising anyone object constructional features region of pedestrian in training sample completely; The so-called image not exclusively comprising pedestrian area, refers to the image only comprising groups of people's object constructional features (such as, the above-mentioned organization of human body of head, hands, foot or part) region.

For finding a kind of mode that pedestrian carries out effectively description, present embodiment is studied based on two hypothesis.

Assume one: the describing mode that the external appearance characteristic (static nature) of pedestrian and motion feature (behavioral characteristics) combine can improve the discriminating power of pedestrian detector; Assume two: for pedestrian, based on the describing mode of gradient relative to, for the describing mode of gray scale, there is better robustness.

Based on the first it is assumed that it is contemplated that pedestrian's external appearance characteristic when transfixion and motion feature when walking, in the present embodiment, pedestrian is expressed as space-time body. Wherein, any one space-time body is all made up of five the positive samples that comprise a group adjacent.

Assume based on the second, in order to take into account external appearance characteristic (static nature) and motion feature (behavioral characteristics), in the present embodiment, the space-time body definition for gradient conversion includes two aspects, is specially as shown in formula (1):

$G\left(p\right)=\mathrm{\&alpha;}\sqrt{{\left(V_x\right)}^2+{\left(V_y\right)}^2}+(1-\mathrm{\&alpha;})\left|V_t\right|---\left(1\right);$

Wherein, V (p) represent on space-time body put p (x, y, gray value t) use V_x��V_yAnd V_tRepresenting along the Grad on x-axis, y-axis and time t axle respectively, G (p) represents gradient conversion space-time body. Section 1 on the right of equation is spatial gradient, it is possible to external appearance characteristic (static nature) is indicated; Section 2 on the right of equation is time gradient, is used for representing that pedestrian's data are along motion feature produced by time shaft (behavioral characteristics). By balance parameters �� to regulate the weight between spatial gradient and time gradient.

In order to be effectively obtained the profile information on space-time body, utilizing range conversion that each pixel on space-time body gives this pixel distance value to its nearest profile point, range conversion space-time body is defined as shown in formula (2):

D (p)=min (dis (p, p^*)), p^*�� p ' | G (p ') > �� } (2);

Wherein, �� represents the threshold value of gradient conversion space-time body, concrete, ��=50 in the present embodiment; Dis () represents Euclidean distance; D (p) represents range conversion space-time body.

Now, so that the gradient of two-value conversion space-time body (i.e. " profile of space-time body ") is converted into range conversion space-time body. In this conversion body, each pixel value represents the current location distance to its nearest profile point. In this way, it is possible to by one-dimensional space profiles curve extension to three-dimensional space-time body.

The appearance information of human body is expressed to obtain pedestrian's movable information in be embodied in multiframe in motor process simultaneously, need on adjacent some two field pictures, extract three-dimensional Haar feature, in the present embodiment, it is carry out extracting three-dimensional Haar feature in adjacent 5 two field pictures (that is: positive sample or negative sample).

Then, perform sub-step S12. conversion space-time body of adjusting the distance and be integrated computing and obtain integration body, and calculate the pixel value sum in integration body.

In order to three-dimensional Haar feature is quickly calculated, present embodiment adopts the computational methods being similar to integrogram, its with prior art in the differring primarily in that of computational methods of integrogram, the present invention carries out the calculating of integrogram on three-dimensional space-time body, and is referred to as integration body. For integration body, it is at (x, y, t) pixel value at place be in space-time body all coordinate figures less than (x, y, t) sum of the pixel value at place, this integration body (x, y, t) shown in the calculation equation below (3) at place:

$\mathrm{IV}(x,y,t)=\underset{x^{\&prime;}<x,y^{\&prime;}<y,t^{\&prime;}<t}{\mathrm{\&Sigma;}}D(x^{\&prime;},y^{\&prime;},t^{\&prime;})---\left(3\right);$

Wherein, (x, y, t) represent integration body to IV, and D (x ', y ', t ') represents range conversion space-time body.

Finally, perform sub-step S13. and utilize three-dimensional Haar wave filter to carry out convolution with integration body, obtain the three-dimensional Haar characteristic vector of space-time body.

In the present embodiment, the three-dimensional Haar wave filter of seven shown in Fig. 2 kind is adopted to extract external appearance characteristic and the motion feature of pedestrian. Specifically, can use static filter S1, S2 and S3 that the external appearance characteristic (static nature) of pedestrian is extracted. Using kinetic filter D1, D2, D3 and D4 to be used for the motion feature of pedestrian is extracted, these wave filter can represent the different motor pattern that pedestrian embodies in motor process.

Kinetic filter D1 and D2 is used for describing pedestrian's produced translational motion in motor process; Kinetic filter D3 is used for describing pedestrian's produced rotary motion in motor process; Kinetic filter D4 is used for the appearing and subsiding describing pedestrian in monitoring region.

Utilize integration body, the pixel value in any one cube body and reducing can be added by 7 steps and complete. Shown in ginseng Fig. 3, the pixel value sum in the cube shown in it can calculate with formula (4) and obtain:

Sum (V)=IV (H)-IV (D)-IV (F)-IV (G)+IV (b)+IV (C)+IV (E)-IV (A) (4);

Wherein, sum (V) represents the pixel value sum in cube A-b-C-D-D-E-F-G-H, then utilizes three-dimensional Haar wave filter to carry out convolution with integration body, it is possible to obtain the three-dimensional Haar characteristic vector on space-time body.

Next perform step S2, based on GentleAdaboost Cascade algorithms, three-dimensional Haar characteristic vector obtained in step S1 be trained, obtain pedestrian's cascade classifier.

Wherein, the positive sample in this positive sample set is the sample (i.e. positive sample) comprising pedestrian head and/or shoulder; The negative sample that this negative sample is concentrated is the sample (i.e. negative sample) not comprising pedestrian head and/or shoulder. Concrete, in the present embodiment, the number of the positive sample in positive sample set in initializing pedestrian's grader is 4000, and the number of the negative sample that negative sample is concentrated is 6000. Owing to GentleAdaboost Cascade algorithms is the routine techniques means of the art, therefore do not repeat them here.

Then, perform step S3, utilize pedestrian's cascade classifier that the input picture obtained from monitoring region is carried out pedestrian detection.

Shown in ginseng Fig. 4, in the present embodiment, video camera 10 vertically shoots and is applicable to outdoor environment and indoor environment. In the present embodiment, " being obtained the video streaming image in monitoring region by video camera " in this step is particularly as follows: by the video streaming image in video camera 10 acquisition monitoring region 30 as input picture, described monitoring region 30 is positioned at the underface of video camera 10.

Concrete, video camera 10 is arranged on the surface of gateway 20, and pedestrian can walk up and down on the direction of arrow 201 in gateway 20. The acquired monitoring region 30 of video camera 10 can be completely covered the Zone Full of gateway 20. This gateway 20 may be provided to be needed in front door or the corridor that the market that pedestrian's number is added up, garage, bank etc. need key monitoring place.

It should be noted that the best results that the present invention is when video camera 10 vertically faces monitoring region 30, certainly also video camera 10 can be faced toward obliquely and need the region carrying out pedestrian's number counting statistics, to cover whole monitoring region 30 by video camera 10.

In the present embodiment, this monitoring region 30 is rectangle; Can certainly be square or circular or other shapes. Video camera 10 is positioned at the surface of the central point 301 in monitoring region 30, and now this monitoring region 30 is positioned at the underface of video camera 10.

Shown in ginseng 5, in the present embodiment, the detection unit of 1 to n level constitute cascade structure, input region to be detected from ground floor detection unit 1, progressively judge whether region to be detected is pedestrian area. If be judged as " non-" in the detection unit of certain level, then this image to be detected is classified as " non-pedestrian region " class; If through the detection unit 1��n of all levels all judge be pedestrian area after, then export pedestrian area. Concrete, it is all utilize the training of GentleAdaboost Cascade algorithms to obtain from the detection unit of 1 to n level.

Finally, perform step S4, based on EKM algorithm the pedestrian detected in step S3 be tracked and count.

EKM algorithm is the method adopting Kalman filter and meanshift algorithm to combine, it is particularly as follows: according to the coordinate of former frame target location in continuous print input picture, utilize Kalman filter to predict the coordinate points that in this frame input picture, target is possible, then utilize meanshift algorithm to be iterated computing with the coordinate points estimated for starting point.

Kalman is the estimation of a kind of recurrence, as long as namely knowing that the estimated value of a upper moment state and the observation of current state just can calculate the estimated value of current state, does not therefore need the historical information of hourly observation or estimation.

The coordinates of targets point estimated due to Kalman filter is compared previous frame target location and is more connect the target location drawing this frame, so when using meanshift algorithm to its this frame of iterative target location, can effectively reduce the number of times of iterative computation, shorten the overall target recognition time. Finally count, obtain the number of pedestrian. Certainly, present embodiment is tracked and counting also by the barycenter of the pedestrian contour detected.

In addition, it is to be understood that, although this specification is been described by according to embodiment, but not each embodiment only comprises an independent technical scheme, this narrating mode of description is only for clarity sake, description should be made as a whole by those skilled in the art, and the technical scheme in each embodiment through appropriately combined, can also form other embodiments that it will be appreciated by those skilled in the art that.

Claims (9)

1. the pedestrian detection method based on space-time body motion outline feature, it is characterised in that comprise the following steps:

S1, the three-dimensional Haar wave filter of use extract the three-dimensional Haar characteristic vector of the space-time body that training sample is concentrated;

S2, based on GentleAdaboost Cascade algorithms, described three-dimensional Haar characteristic vector is trained, obtains pedestrian's cascade classifier;

S3, utilize pedestrian's cascade classifier to from monitoring region obtain input picture carry out pedestrian detection;

S4, based on EKM algorithm the pedestrian detected in step S3 it is tracked and counts.

2. pedestrian detection method according to claim 1, it is characterized in that, described training sample set includes positive sample set and negative sample collection, described positive sample set is made up of some positive samples comprising pedestrian area image, and described negative sample collection is made up of some negative samples not comprising and/or not exclusively comprising pedestrian area image.

3. pedestrian detection method according to claim 1 and 2, it is characterised in that described step S1 particularly as follows:

S11. the 5 adjacent positive samples of frame or negative sample are defined as space-time body go forward side by side row space and time gradient conversion, so that the profile of space-time body is converted into range conversion space-time body;

S12. conversion space-time body of adjusting the distance is integrated computing and obtains integration body, and calculates the pixel value sum in integration body;

S13. utilize three-dimensional Haar wave filter to carry out convolution with integration body, obtain the three-dimensional Haar characteristic vector of space-time body.

4. pedestrian detection method according to claim 3, it is characterised in that positive sample in described step S11 or negative sample are pixel size be 30 �� 30 256 rank gray level images.

5. pedestrian detection method according to claim 3, it is characterised in that the integral operation formula in described step S12 is:

Wherein, (x, y, t) for integration body, D (x ', y ', t ') is range conversion space-time body to IV.

6. pedestrian detection method according to claim 5, it is characterized in that, the region of described integration body is the space structure being included in range conversion space-time body, described integration body is at (x, y, t) the pixel value sum at place is that in space-time body, all coordinate figures are respectively less than (x, y, t) the pixel value sum at place.

7. pedestrian detection method according to claim 3, it is characterized in that, three-dimensional Haar wave filter in described step S13 is by being used for extracting three static Haar wave filter S1, S2, S3 of pedestrian's static exercise feature, and four dynamic Haar filter Ds 1, D2, D3, D4 for extracting pedestrian's behavioral characteristics collectively constitute.

8. pedestrian detection method according to claim 1, it is characterised in that described " from the input picture that monitoring region is acquired " in step S3 is the video streaming image being obtained monitoring region by video camera, and described monitoring region is positioned at the underface of video camera.

9. pedestrian detection method according to claim 1, it is characterized in that, described " EKM algorithm " in step S4 is particularly as follows: according to the coordinate of former frame target location in continuous print input picture, utilize Kalman filter to predict the coordinate points that in this frame input picture, target is possible, then utilize meanshift algorithm to be iterated computing with the coordinate points estimated for starting point.