CN104899559A - Rapid pedestrian detection method based on video monitoring - Google Patents

Abstract

The present invention relates to a rapid pedestrian detection method based on video monitoring. The detection method comprises: carrying out image sampling; carrying out parameter setting and image preprocessing; carrying out motion detection; carrying out pedestrian detection by using a CENTRIST descriptor and a linear SVM classifier; and the like. With the rapid pedestrian detection method based on video monitoring according to the present invention, cost in the actual monitoring process is effectively reduced, and real-time monitoring on a video and rapid and accurate pedestrian detection are implemented.

Description

A kind of rapid pedestrian detection method based on video monitoring

Technical field

The invention belongs to field of intelligent video surveillance, be specifically related to a kind of rapid pedestrian detection method based on video monitoring.

Background technology

Be used as the characteristics of image of pedestrian detection from small echo first time and use SVM algorithm to target decision classification of taking exercises, to Frenchman Dalal in 2005 when CVPR delivers the detection algorithm of HOG+SVM, the basic frame structure of pedestrian detection algorithm just establishes, namely first by choosing suitable image characteristics extraction image information, set up suitable disaggregated model training standard image pattern, and then obtaining rational recognition classifier for image to be detected, final detection obtains pedestrian position and size.After this, pedestrian detection algorithm development is more and more faster, the feature extracted is also no longer single, but the edge contour from target image of suiting measures to local conditions, texture information are to histogram of gradients, colouring information etc., and the sorter adapted with it also has neural network, SVM, Adaboost and degree of depth study scheduling algorithm.

Classical HOG+SVM algorithm, can find 2 deficiencies by experiment: it is oversize that (1) image detects the time consumed, cannot reach the rate request for photographic technique in practical application far away; (2) the sample set not collecting corresponding scene, do not train corresponding categorised decision function prerequisite under, the effect of actual pedestrian detection is quite inaccurate, crosses inspection, undetected situation is very serious.This shows, although the training set of the SVM decision function acquiescence carried in OpenCV is INRIA standard pedestrian image library, Detection results is unsatisfactory.In the video monitoring of reality, in order to improve accuracy, must for different testing environments, scene, the corresponding standard picture storehouse of dynamic collection, and calculate corresponding decision function.

For first deficiency, (namely Census Transform Histogram, also claims C to a kind of CENTRIST descriptor being similar to HOG descriptor that the present invention proposes ⁴), its computation complexity on the basis of same projecting edge information but reduces greatly.Compared to HOG descriptor, C ⁴more pay attention to the information such as the edge contour of reservation image, effectively capture the local feature of image and complete coding, avoiding because illumination etc. changes suddenly the harmful effect caused.And it needs the mathematic interpolation amount calculated between neighborhood territory pixel to be greatly reduced.

In order to improve accuracy and the travelling speed of pedestrian detection, the present invention proposes a kind of innovatory algorithm on original svm classifier algorithm basis---the algorithm of support vector machine (Fast Classification Support Vector Machine, is abbreviated as: FCSVM) of Fast Classification.Experiment proves that this algorithm reduces the number of the support vector of training out on the basis of not obvious reduction SVM algorithm classification effect, also can decrease the time overhead of decision function during detection accordingly.

Summary of the invention

For above-mentioned conventional pedestrian's detection method exist problem, the present invention proposes a kind of rapid pedestrian detection method based on video monitoring, can improve accuracy and the travelling speed of pedestrian detection.

For achieving the above object, the concrete technical scheme of the present invention is as follows: a kind of rapid pedestrian detection method based on video monitoring, comprises the steps:

1) video acquisition and setting parameter: read video flowing, initial threshold value Thread, goes to step 2);

2) motion detection counter i=0, goes to step 3);

3) read a two field picture, go to step 4);

4) judge whether counter i is greater than threshold value Thread, if i>Thread, go to step 8), otherwise go to step 5);

5) Image semantic classification: carry out colour space transformation, denoising and gray processing process to image, utilizes Gamma formula compressed image size, goes to step 6);

6) moving target has been judged whether: detect in video image and whether have moving target to swarm into, if do not have moving target, go to step 2); Otherwise go to step 7);

7) counter i adds 1:i=i+1, goes to step 3);

8) pedestrian detection: judge whether the moving target swarmed into is pedestrian, if pedestrian goes to step 9), otherwise go to step 2);

9) report to the police;

Wherein, above-mentioned steps 8) comprise following steps:

8.1) the CENTRIST descriptor of present frame is extracted;

8.2) pedestrian's unique point in the Linear SVM detection of classifier image trained is used;

8.3) judge whether the moving object swarmed into is pedestrian according to pedestrian's unique point.

Further, above-mentioned steps 8.1) middle extraction CENTRIST descriptor, comprise the steps:

8.1.1) carry out gray processing, smothing filtering to present frame I, eliminate Local textural feature, extract the magnitude relationship between basic marginal information and pixel, wherein, wave filter is Sobel operator;

8.1.2) according to the magnitude relationship between present frame I pixel, a new image I is constructed ^/;

8.1.3) image I is asked ^/in the CT encoded radio of each pixel;

8.1.4) according to image I ^/in the histogram of CT encoded radio of all pixels obtain CENTRIST descriptor.

Further, above-mentioned steps 8.2) in use the method for pedestrian's unique point in the Linear SVM detection of classifier image that trains to comprise the steps:

8.2.1) carry out SVM training to the sample data collection gathered, obtain support vector collection S, wherein, the dimension of vector set S is N;

8.2.2) m=1 is established, n=N-m;

8.2.3) according to m, n, vector set S is divided into two subset P and Q, wherein, the dimension of subset P is the dimension of m, subset Q is n;

8.2.4) the inner product Km of subset of computations P and the inner product Kn of subset Q, computing formula is as follows:

$K_m=\left[\begin{array}{ccc}\mathrm{\φ}\left(x_1\right)\mathrm{\φ}\left(s_1\right)& ......& \mathrm{\φ}\left(x_L\right)\mathrm{\φ}\left(s_1\right)\\ \·& & \·\\ \·& ......& \·\\ \·& & \·\\ \mathrm{\φ}\left(x_1\right)\mathrm{\φ}\left(s_m\right)& ......& \mathrm{\φ}\left(x_L\right)\mathrm{\φ}\left(s_m\right)\end{array}\right]$

$K_n=\left[\begin{array}{ccc}\mathrm{\φ}\left(x_1\right)\mathrm{\φ}\left(s_{m+1}\right)& ......& \mathrm{\φ}\left(x_L\right)\mathrm{\φ}\left(s_{m+1}\right)\\ \·& & \·\\ \·& ......& \·\\ \·& & \·\\ \mathrm{\φ}\left(x_1\right)\mathrm{\φ}\left(s_{N-m}\right)& ......& \mathrm{\φ}\left(x_L\right)\mathrm{\φ}\left(s_{N-m}\right)\end{array}\right]$

8.2.5) transformation matrix W is calculated according to inner product Km, Kn ^t, computing formula is as follows:

K _n＝W ^TK _m；

8.2.6) according to transformation matrix W ^tcompute matrix A and matrix B, computing method are as follows:

A＝(θ ₁,…,θ _L)；

B＝(θ′ ₁,…θ′ _L)

Wherein:

${\mathrm{\θ}}_i={\∑}_{j=1}^N{a}_j{y}_{j}K(x_i,s_j)$

${\mathrm{\θ}}_i^{\′}=(A_m^T+A_n^T{W}^T)K_m$

${\mathrm{\Σ}}_{j=1}^N{a}_j{y}_{j}K(x,s_j)=(A_m^T+A_n^T{W}^T)K_t$

$A_m^T=(a_1{y}_1,a_2{y}_2,...,a_m{y}_m)$

$A_n^T=(a_{m+1}{y}_{m+1},a_{m+2}{y}_{m+2},...,a_N{y}_N)$

K _t＝(K _t.1,…,K _t.m) ^T

In above formula: a _jfor Lagrange multiplier, y _jfor positive and negative mark

8.2.7) ε is asked according to matrix A and matrix B _m=|| A-B||;

8.2.8)m＝m+1，n＝N-m；

8.2.9) judge whether m is less than N, if m<N, then go to step 8.2.3), otherwise go to step 8.2.10);

8.2.10) obtain minimum ε, obtain the transformation matrix W of its correspondence ^t;

8.2.11) at previous step W ^tbasis on, in image, whether pixel is pedestrian's unique point, and wherein decision function is as follows to utilize decision function to judge:

$f\left(x\right)=\mathrm{sgn}((A_m^T+A_n^T{W}^T)K_t+b),$ In formula, b is threshold value (b=y _j-W ^tx);

A kind of rapid pedestrian detection method based on video monitoring that the present invention proposes, effectively reduces the expense in actual monitored, realizes photographic technique and pedestrian detection fast and accurately.

Accompanying drawing explanation

Fig. 1 is based on the rapid pedestrian detection method process flow diagram of video monitoring.

Fig. 2 pedestrian detection process flow diagram.

Fig. 3 extracts CENTRIST descriptor process flow diagram.

Pedestrian's unique point algorithm flow chart in Fig. 4 Linear SVM detection of classifier image.

Embodiment

Below in conjunction with the drawings and specific embodiments, the invention will be further described.

Figure 1 shows that the rapid pedestrian detection method process flow diagram based on video monitoring, comprise the following steps:

1) video acquisition and setting parameter: read video flowing, initial threshold value Thread, in the present embodiment, initialization Thread is 3; Go to step 2);

2) establish motion detection counter i=0, go to step 3);

3) read a two field picture, go to step 4);

4) judge whether counter i is greater than threshold value Thread, if i>Thread, go to step 8), otherwise go to step 5);

5) Image semantic classification: carry out colour space transformation, denoising and gray processing process to image, utilizes Gamma formula compressed image size, goes to step 6);

6) moving target has been judged whether: detect in video image and whether have moving object to swarm into, if do not have moving target, go to step 2); Otherwise go to step 7);

7) counter i adds 1:i=i+1, goes to step 3);

8) pedestrian detection: judge whether moving object is pedestrian, if pedestrian goes to step 9), otherwise go to step 2);

9) report to the police.

Figure 2 shows that pedestrian detection process flow diagram, comprise following steps:

8.1) the CENTRIST descriptor of present frame is extracted;

8.2) pedestrian's unique point in the Linear SVM detection of classifier image trained is used;

8.3) judge whether the moving object swarmed into is pedestrian according to pedestrian's unique point.

Figure 3 shows that and extract CENTRIST descriptor process flow diagram, comprise the steps:

8.1.1) carry out gray processing, smothing filtering to present frame I, eliminate Local textural feature, extract the magnitude relationship between basic marginal information and pixel, wherein, wave filter is Sobel operator;

8.1.2) according to the magnitude relationship between present frame I pixel, a new image I is constructed ^/; In the present embodiment, building method is: according to the magnitude relationship between image I pixel, constructs a new images I ^/still retain the pixel size comparison signal of source images I, and ignore the value size of pixel, namely for any one adjacent pixel to meeting following relation:

sgn(I(p ₁)-I(p ₂))＝sgn(I'(p ₁)-I'(p ₂))

Wherein p1 and p2 is arbitrary neighborhood point;

8.1.3) image I is asked ^/in the CT encoded radio of each pixel;

8.1.4) according to image I ^/in the histogram of CT encoded radio of all pixels obtain CENTRIST descriptor.

Figure 4 shows that pedestrian's unique point algorithm flow chart in Linear SVM detection of classifier image, comprise the steps:

8.2.1) carry out SVM training to the sample data collection gathered, obtain support vector collection S, wherein, the dimension of vector set S is N;

8.2.2) m=1 is established, n=N-m;

8.2.3) according to m, n, vector set S is divided into two subset P and Q, wherein, the dimension of subset P is the dimension of m, subset Q is n;

8.2.4) the inner product Km of subset of computations P and the inner product Kn of subset Q, computing formula is as follows:

$K_m=\left[\begin{array}{ccc}\mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_1\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_1\right)\\ \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& ......& \&CenterDot;\\ \&CenterDot;& & \&CenterDot;\\ \mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_m\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_m\right)\end{array}\right]$

$K_n=\left[\begin{array}{ccc}\mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_{m+1}\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_{m+1}\right)\\ \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& ......& \&CenterDot;\\ \&CenterDot;& & \&CenterDot;\\ \mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_{N-m}\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_{N-m}\right)\end{array}\right]$

8.2.5) transformation matrix W is calculated according to inner product Km, Kn ^t, computing formula is as follows:

K _n＝W ^TK _m；

8.2.6) according to transformation matrix W ^tcompute matrix A and matrix B, computing method are as follows:

A＝(θ ₁,…,θ _L)；

B＝(θ′ ₁,…θ′ _L)

Wherein:

${\mathrm{\&theta;}}_i={\&Sum;}_{j=1}^N{a}_j{y}_{j}K(x_i,s_j)$

${\mathrm{\&theta;}}_i^{\&prime;}=(A_m^T+A_n^T{W}^T)K_m$

${\mathrm{\&Sigma;}}_{j=1}^N{a}_j{y}_{j}K(x,s_j)=(A_m^T+A_n^T{W}^T)K_t$

$A_m^T=(a_1{y}_1,a_2{y}_2,...,a_m{y}_m)$

$A_n^T=(a_{m+1}{y}_{m+1},a_{m+2}{y}_{m+2},...,a_N{y}_N)$

K _t＝(K _t.1,…,K _t.m) ^T

In above formula: a _jfor Lagrange multiplier, y _jfor positive and negative mark

8.2.7) ε is asked according to matrix A and matrix B _m=|| A-B||;

8.2.8)m＝m+1，n＝N-m；

8.2.9) judge whether m is less than N, if m<N, then go to step 8.2.3), otherwise go to step 8.2.10);

8.2.10) obtain minimum ε, obtain the transformation matrix W of its correspondence ^t;

8.2.11) at previous step W ^tbasis on, in image, whether pixel is pedestrian's unique point, and wherein decision function is as follows to utilize decision function to judge:

$f\left(x\right)=\mathrm{sgn}((A_m^T+A_n^T{W}^T)K_t+b),$ In formula, b is threshold value (b=y _j-W ^tx).

Claims (3)

1., based on a rapid pedestrian detection method for video monitoring, it is characterized in that comprising the steps:

1) video acquisition and setting parameter: read video flowing, initial threshold value Thread, goes to step 2);

2) motion detection counter i=0, goes to step 3);

3) read a two field picture, go to step 4);

4) judge whether counter i is greater than threshold value Thread, if i>Thread, go to step 8), otherwise go to step 5);

5) Image semantic classification: carry out colour space transformation, denoising and gray processing process to image, utilizes Gamma formula compressed image size, goes to step 6);

6) moving target has been judged whether: detect in video image and whether have moving target to swarm into, if do not have moving target, go to step 2); Otherwise go to step 7);

7) counter i adds 1:i=i+1, goes to step 3);

8) pedestrian detection: judge whether moving target is pedestrian; If pedestrian goes to step 9), otherwise go to step 2);

9) report to the police;

Wherein, described step 8) comprise following steps:

8.1) the CENTRIST descriptor of present frame is extracted;

8.2) pedestrian's unique point in the Linear SVM detection of classifier image trained is used;

8.3) judge whether the moving target swarmed into is pedestrian according to pedestrian's unique point.

2., as claimed in claim 1 based on the rapid pedestrian detection method of video monitoring, it is characterized in that: described step 8.1) middle extraction CENTRIST descriptor, comprise the steps:

8.1.1) carry out gray processing, smothing filtering to present frame I, eliminate Local textural feature, extract the magnitude relationship between basic marginal information and pixel, wherein, wave filter is Sobel operator;

8.1.2) according to the magnitude relationship between present frame I pixel, a new image I is constructed ^/;

8.1.3) image I is asked ^/in the CT encoded radio of each pixel;

8.1.4) according to image I ^/in the histogram of CT encoded radio of all pixels obtain CENTRIST descriptor.

3., as claimed in claim 1 based on the rapid pedestrian detection method of video monitoring, it is characterized in that: described step 8.2) in use the method for pedestrian's unique point in the Linear SVM detection of classifier image trained to comprise the steps:

8.2.1) carry out SVM training to the sample data collection gathered, obtain support vector collection S, wherein, the dimension of vector set S is N;

8.2.2) m=1 is established, n=N-m;

8.2.3) according to m, n, vector set S is divided into two subset P and Q, wherein, the dimension of subset P is the dimension of m, subset Q is n;

8.2.4) the inner product Km of subset of computations P and the inner product Kn of subset Q, computing formula is as follows:

$K_m=\left[\begin{array}{ccc}\mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_1\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_1\right)\\ \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& ......& \&CenterDot;\\ \&CenterDot;& & \&CenterDot;\\ \mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_m\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_m\right)\end{array}\right]$

$K_n=\left[\begin{array}{ccc}\mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_{m+1}\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_{m+1}\right)\\ \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& ......& \&CenterDot;\\ \&CenterDot;& & \&CenterDot;\\ \mathrm{\&phi;}\left(x_1\right)\mathrm{\&phi;}\left(s_{N-m}\right)& ......& \mathrm{\&phi;}\left(x_L\right)\mathrm{\&phi;}\left(s_{N-m}\right)\end{array}\right]$

8.2.5) transformation matrix W is calculated according to inner product Km, Kn ^t, computing formula is as follows:

K _n＝W ^TK _m；

8.2.6) according to transformation matrix W ^tcompute matrix A and matrix B, computing method are as follows:

A＝(θ ₁,…,θ _L)；

B＝(θ′ ₁,…θ′ _L)

Wherein:

${\mathrm{\&theta;}}_i={\mathrm{\&Sigma;}}_{j=1}^N{a}_j{y}_{j}K(x_i,s_j)$

${\mathrm{\&theta;}}_i^{\&prime;}=(A_m^T+A_n^T{W}^T)K_m$

${\mathrm{\&Sigma;}}_{j=1}^N{a}_j{y}_{j}K(x,s_j)=(A_m^T+A_n^T{W}^T)K_t$

$A_m^T=(a_1{y}_1,a_2{y}_2,...,a_m{y}_m)$

$A_n^T=(a_{m+1}{y}_{m+1},a_{m+2}{y}_{m+2},...,a_N{y}_N)$

K _t＝(K _t.1,…,K _t.m) ^T

In above formula: a _jfor Lagrange multiplier, y _jfor positive and negative mark

8.2.7) ε is asked according to matrix A and matrix B _m=|| A-B||;

8.2.8)m＝m+1,n＝N-m；

8.2.9) judge whether m is less than N, if m<N, then go to step 8.2.3), otherwise go to step 8.2.10);

8.2.10) obtain minimum ε, obtain the transformation matrix W of its correspondence ^t;

8.2.11) at previous step W ^tbasis on, in image, whether pixel is pedestrian's unique point, and wherein decision function is as follows to utilize decision function to judge:

$f\left(x\right)=\mathrm{sgn}((A_m^T+A_n^T{W}^T)K_t+b),$ In formula, b is threshold value (b=y _j-W ^tx).