CN105187801A - Condensed video generation system and method - Google Patents

Abstract

The invention relates to a condensed video generation system and method. The system comprises a detection recognition tracking and extracting module, an optimizing and combining module, a video frame background module, a trajectory background pasting module, and a video condensing module. The detection recognition tracking and extracting module detects and recognizes the moving objects within a moving targeted area, tracks the detected and recognized moving objects and extracts the trajectories of the detected and recognized moving objects. The optimizing and combining module optimizes and combines the extracted moving trajectories of the moving objects. The video frame background module dynamically updates the video background of the moving targeted area. The trajectory background pasting module combines the moving objected with the updated video background in a matched pasting manner according to the combined and optimized moving trajectories in order to generate a condensed video frame. The video condensing module generates a condensed video according to the condensed video frame. In comparison with the prior art, the condensed video formed by the invention is shorter than an original video and most of information concerning the moving objects are still kept in the condensed video so that people can conveniently browse and the system and method are high in processing efficiency.

Description

A kind of generation system of summarized radio and method

Technical field

The present invention relates to field of video image processing, particularly relate to a kind of generation system and method for summarized radio.

Background technology

Along with the develop rapidly of monitoring technique and network technology, high-definition network monitoring camera is widely used in industry-by-industry.These high-definition camera all weather operations, create the monitor video data of magnanimity, but, because the limitation of monitoring human resources and time resource usually causes some massive video data from not processed.How long monitor video of browsing complete fast has become the current problem demanding prompt solution of monitoring trade; The important means that summarized radio solves exactly " massive video data process ".But, the traditional summarized radio generation method based on key frame adopt frame sampling cannot each object of complete representation movement locus and result in the loss of a large amount of useful video information.

Summary of the invention

Technical problem to be solved by this invention is to provide a kind of summarized radio shorter than original video, remains the information of most of Moving Objects in original video, is convenient to fast browsing, and the generation system of the high summarized radio for the treatment of effeciency and method.

The technical scheme that the present invention solves the problems of the technologies described above is as follows: a kind of generation system of summarized radio, comprises and detects recognition and tracking extraction module, Combinatorial Optimization module, frame of video background module, track background stickup module, summarized radio module;

Described detection recognition and tracking extraction module, for detecting the Moving Objects identified in motion target area, follows the tracks of and movement locus extraction detecting the Moving Objects identified;

Described Combinatorial Optimization module, for carrying out Combinatorial Optimization to the movement locus extracted;

Described frame of video background module, for dynamically updating the video background in motion target area;

Described track background stickup module, for according to the movement locus after Combinatorial Optimization by Moving Objects with upgrade after video background carry out mate pasting and merge, generation summarized radio frame;

Described summarized radio module, for generating summarized radio according to summarized radio frame.

The invention has the beneficial effects as follows: by detecting the running of recognition and tracking extraction module, Combinatorial Optimization module, frame of video background module, track background stickup module and summarized radio module coordination, generate the summarized radio shorter than original video, remain the information of most of Moving Objects in original video, be convenient to fast browsing, and treatment effeciency is high.

On the basis of technique scheme, the present invention can also do following improvement.

Further, described detection recognition and tracking extraction module identifies the Moving Objects in motion target area by setting up rank adaptive model; Described rank adaptive model is with the HOG (HistogramofGradient of cart headstock and dolly vehicle body image, finger direction histogram of gradients) feature is as Expressive Features, and in conjunction with Linear SVM (SupportVectorMachine, referring to SVMs) full features training sorting technique carries out the detection identification of vehicle vehicle, simultaneously, target vehicle region was extracted roughly by GMM (GaussianMixtureModel refers to gauss hybrid models) before HOG feature extraction.

The beneficial effect of above-mentioned further scheme is adopted to be: to establish rank adaptive model, ensure that, on the Demand Base that people and Che separately browse, the accuracy of detection and speed have all had larger lifting, and taken into account the robustness of reply scene changes, reduce operand, improve precision.

Further, the full features training sorting technique of described SVM is specially:

Can assign to say for linear two classes, hyperplane, i.e. a linear classifier be found, all training samples correctly can be classified, namely meet:

$y_i\[w^T{x}_i+b\]>0,\∀i=1,...,n$

Wherein

Yi represents the class label giving input data, and xi represents i-th input HOG characteristic vector;

Due to the nonuniqueness of hyperplane, solve optimal hyperlane by the method for quadratic programming;

$\left\{\begin{array}{c}{w}^{*}=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i{y}_i{x}_i\\ b^{*}=y_i-w^{*T}{x}_i\end{array}\right.$

Wherein, α _irepresent fixing Lagrange multiplier, w ^*represent the normal vector of optimal hyperlane, b ^*represent the side-play amount of optimal hyperlane.

Further, described detection recognition and tracking extraction module carries out moving object tracking by multiple-camera method for tracking target, and described multiple-camera method for tracking target is specially:

$\left\{\begin{array}{c}{p}_{x_1}^{camA/t}(x,y)=H_{camB-camA}\×Foo{t}_A^{camB/t}(x,y)\\ p_{x_2}^{camA/t}(x,y)=H_{camB-camA}\×Foo{t}_B^{camB/t}(x,y)\end{array}\right.$

$\left\{\begin{array}{c}{p}_{x_1}^{camA/t}(x,y)\∈Area\left(x\right)\\ p_{x_2}^{camA/t}(x,y)\∈Area\left(x\right)\end{array}\right.$

Wherein represent t, the coordinate position of minimum point coordinate in camB of target A, H _camB-camAthe affine projection matrix of camB to camA scene, represent the coordinate of any point in t camA, Area (x) is ground regional extent shared by target.

The beneficial effect of above-mentioned further scheme is adopted to be: multiple-camera method for tracking target solves target and blocks or be separated the failed problem of rear tracking, has higher robustness; Multiple situations is there is in the same target of situation eliminating space-time confusion in a summarized radio picture.

Further, the video background of described frame of video background module dynamically updates and is specially: vote to video background at each Moving Objects track of a frame of video, selects the background of all Moving Objects tracks under original video environment.

Further, the video background after renewal and movement locus agglomerate background are carried out the pixel fusion algorithm process of Gaussian Profile by described track background stickup module.

Further, the image pixel blending algorithm of described Gaussian Profile is specially:

F(m,n)＝ω ₁P _MV(m,n)+ω ₂P _BG(m,n)

In above formula, P _mV(m, n) and P _bG(m, n) is respectively the pixel value of the image BG of each Moving Objects MV in summarized radio frame and frame background of making a summary; M and n is respectively line number and the row number of pixel in image, and m=1,2 ..., M, n=1,2 ..., N; w ₁, w ₂for weight coefficient, and w ₁+ w ₂=1;

${\mathrm{Corr}}_{(BG,MV)}=\frac{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}{\sqrt{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]}^2\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}^2}}$

In formula, Corr is the coefficient correlation of MV and BG two width source images, for the pixel average of source images MV, for the pixel average of source images BG;

$\left\{\begin{array}{c}{\mathrm{\ω}}_1=\frac{1}{2}(1-\left|Corr\right|)\\ {\mathrm{\ω}}_2=1-{\mathrm{\ω}}_1\end{array}\right.$

$G(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{{(x-x_c)}^2+{(y-y_c)}^2}{2{\mathrm{\σ}}^2}}$

Wherein, σ is the standard deviation of normal distribution, (x _c, y _c) for Gaussian function distribution barycenter.

The beneficial effect of above-mentioned further scheme is adopted to be: the image pixel blending algorithm of Gaussian Profile can eliminate splicing vestige, realizes the seamless fusion of the movement locus of summarized radio background and tracking; Retain the information of original image, eliminate deceptive information, obtain reliable and stable syncretizing effect.

A generation method for summarized radio, comprises the following steps:

Step S1. detects the Moving Objects identified in motion target area, follows the tracks of and movement locus extraction detecting the Moving Objects identified;

Step S2. carries out Combinatorial Optimization to the movement locus extracted; Video background in motion target area is dynamically updated;

Step S3. according to the movement locus after Combinatorial Optimization by Moving Objects with upgrade after video background carry out mate pasting and merge, generation summarized radio frame;

Step S4. generates summarized radio according to summarized radio frame.

Further, the specific implementation of the Moving Objects identified in motion target area is detected in described step S1: by setting up rank adaptive model, the Moving Objects in motion target area is identified; Described rank adaptive model is using the HOG feature of cart headstock and dolly vehicle body image as Expressive Features, and the detection identification of vehicle vehicle is carried out in conjunction with the full features training sorting technique of Linear SVM, meanwhile, before HOG feature extraction, target vehicle region is extracted roughly by GMM.

Further, the specific implementation of described step S3: the pixel fusion algorithm process of the video background after renewal and movement locus agglomerate background being carried out Gaussian Profile; The image pixel blending algorithm of described Gaussian Profile is specially:

F(m,n)＝ω ₁P _MV(m,n)+ω ₂P _BG(m,n)

In above formula, P _mV(m, n) and P _bG(m, n) is respectively the pixel value of the image BG of each Moving Objects MV in summarized radio frame and frame background of making a summary; M and n is respectively line number and the row number of pixel in image, and m=1,2 ..., M, n=1,2 ..., N; w ₁, w ₂for weight coefficient, and w ₁+ w ₂=1;

${\mathrm{Corr}}_{(BG,MV)}=\frac{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}{\sqrt{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]}^2\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}^2}}$

In formula, Corr is the coefficient correlation of MV and BG two width source images, for the pixel average of source images MV, for the pixel average of source images BG;

$\left\{\begin{array}{c}{\mathrm{\ω}}_1=\frac{1}{2}(1-\left|Corr\right|)\\ {\mathrm{\ω}}_2=1-{\mathrm{\ω}}_1\end{array}\right.$

$G(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{{(x-x_c)}^2+{(y-y_c)}^2}{2{\mathrm{\σ}}^2}}$

Wherein, σ is the standard deviation of normal distribution, (x _c, y _c) for Gaussian function distribution barycenter.

The invention has the beneficial effects as follows: the inventive method proposes a kind of target detection recognition methods based on rank adaptive model, and propose a kind of multiple-camera method for tracking target based on event analysis, in the process that monitoring summarized radio generates, by voting in the best background of all Moving Objects tracks under original video environment that can reflect and participate in merging, carry out dynamically updating of summarized radio background; Meanwhile, adopt rectangle Gauss weight allocation strategy, well achieve the seamless pixel fusion effect of summarized radio frame background and Moving Objects track.

Accompanying drawing explanation

Fig. 1 is the module frame chart of the generation system of a kind of summarized radio of the present invention;

Fig. 2 is the flow chart of the generation method of a kind of summarized radio of the present invention;

Fig. 3 is the algorithm of target detection schematic diagram based on rank adaptive model;

Fig. 4 is that two video cameras carry out target tracking algorism flow chart;

Fig. 5 is sdi video target profile;

Fig. 6 is sdi video goal programming distribution map.

In accompanying drawing, the list of parts representated by each label is as follows:

1, recognition and tracking extraction module is detected, 2, Combinatorial Optimization module, 3, frame of video background module, 4, track background stickup module, 5, summarized radio module.

Embodiment

Be described principle of the present invention and feature below in conjunction with accompanying drawing, example, only for explaining the present invention, is not intended to limit scope of the present invention.

As shown in Figure 1, a kind of generation system of summarized radio, comprises and detects recognition and tracking extraction module 1, Combinatorial Optimization module 2, frame of video background module 3, track background stickup module 4, summarized radio module 5;

Described detection recognition and tracking extraction module 1, for detecting the Moving Objects identified in motion target area, follows the tracks of and movement locus extraction detecting the Moving Objects identified;

Described Combinatorial Optimization module 2, for carrying out Combinatorial Optimization to the movement locus extracted;

Described frame of video background module 3, for dynamically updating the video background in motion target area;

Described track background stickup module 4, for according to the movement locus after Combinatorial Optimization by Moving Objects with upgrade after video background carry out mate pasting and merge, generation summarized radio frame;

Described summarized radio module 5, for generating summarized radio according to summarized radio frame.

Preferably, described detection recognition and tracking extraction module 1 identifies the Moving Objects in motion target area by setting up rank adaptive model; Described rank adaptive model is using the HOG feature of cart headstock and dolly vehicle body image as Expressive Features, and the detection identification of vehicle vehicle is carried out in conjunction with the full features training sorting technique of Linear SVM, meanwhile, before HOG feature extraction, target vehicle region is extracted roughly by GMM.

Preferably, the full features training sorting technique of described SVM is specially:

Can assign to say for linear two classes, hyperplane, i.e. a linear classifier be found, all training samples correctly can be classified, namely meet:

$y_i\[w^T{x}_i+b\]>0,\∀i=1,...,n$

Wherein

Yi represents the class label giving input data, and xi represents i-th input HOG characteristic vector;

Due to the nonuniqueness of hyperplane, solve optimal hyperlane by the method for quadratic programming;

$\left\{\begin{array}{c}{w}^{*}=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i{y}_i{x}_i\\ b^{*}=y_i-w^{*T}{x}_i\end{array}\right.$

Wherein, α _irepresent fixing Lagrange multiplier, w ^*represent the normal vector of optimal hyperlane, b ^*represent the side-play amount of optimal hyperlane.

Preferably, described detection recognition and tracking extraction module 1 carries out moving object tracking by multiple-camera method for tracking target, and described multiple-camera method for tracking target is specially:

$\left\{\begin{array}{c}{p}_{x_1}^{camA/t}(x,y)=H_{camB-camA}\×Foo{t}_A^{camB/t}(x,y)\\ p_{x_2}^{camA/t}(x,y)=H_{camB-camA}\×Foo{t}_B^{camB/t}(x,y)\end{array}\right.$

$\left\{\begin{array}{c}{p}_{x_1}^{camA/t}(x,y)\∈Area\left(x\right)\\ p_{x_2}^{camA/t}(x,y)\∈Area\left(x\right)\end{array}\right.$

Wherein represent t, the coordinate position of minimum point coordinate in camB of target A, H _camB-camAthe affine projection matrix of camB to camA scene, represent the coordinate of any point in t camA, Area (x) is ground regional extent shared by target.

Preferably, the video background of described frame of video background module 3 dynamically updates and is specially: vote to video background at each Moving Objects track of a frame of video, selects the background of all Moving Objects tracks under original video environment.

Preferably, the video background after renewal and movement locus agglomerate background are carried out the pixel fusion algorithm process of Gaussian Profile by described track background stickup module 4.

Preferably, the image pixel blending algorithm of described Gaussian Profile is specially:

F(m,n)＝ω ₁P _MV(m,n)+ω ₂P _BG(m,n)

In above formula, P _mV(m, n) and P _bG(m, n) is respectively the pixel value of the image BG of each Moving Objects MV in summarized radio frame and frame background of making a summary; M and n is respectively line number and the row number of pixel in image, and m=1,2 ..., M, n=1,2 ..., N; w ₁, w ₂for weight coefficient, and w ₁+ w ₂=1;

${\mathrm{Corr}}_{(BG,MV)}=\frac{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}{\sqrt{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]}^2\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}^2}}$

In formula, Corr is the coefficient correlation of MV and BG two width source images, for the pixel average of source images MV, for the pixel average of source images BG;

$\left\{\begin{array}{c}{\mathrm{\ω}}_1=\frac{1}{2}(1-\left|Corr\right|)\\ {\mathrm{\ω}}_2=1-{\mathrm{\ω}}_1\end{array}\right.$

$G(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{{(x-x_c)}^2+{(y-y_c)}^2}{2{\mathrm{\σ}}^2}}$

Wherein, σ is the standard deviation of normal distribution, (x _c, y _c) for Gaussian function distribution barycenter.

As shown in Figure 2, a kind of generation method of summarized radio, comprises the following steps:

Step S1. detects the Moving Objects identified in motion target area, follows the tracks of and movement locus extraction detecting the Moving Objects identified;

Step S2. carries out Combinatorial Optimization to the movement locus extracted; Video background in motion target area is dynamically updated;

Step S3. according to the movement locus after Combinatorial Optimization by Moving Objects with upgrade after video background carry out mate pasting and merge, generation summarized radio frame;

Step S4. generates summarized radio according to summarized radio frame.

Preferably, the specific implementation of the Moving Objects identified in motion target area is detected in described step S1: by setting up rank adaptive model, the Moving Objects in motion target area is identified; Described rank adaptive model is using the HOG feature of cart headstock and dolly vehicle body image as Expressive Features, and the detection identification of vehicle vehicle is carried out in conjunction with the full features training sorting technique of Linear SVM, meanwhile, before HOG feature extraction, target vehicle region is extracted roughly by GMM.

Preferably, the specific implementation of described step S3: the pixel fusion algorithm process of the video background after renewal and movement locus agglomerate background being carried out Gaussian Profile; The image pixel blending algorithm of described Gaussian Profile is specially:

F(m,n)＝ω ₁P _MV(m,n)+ω ₂P _BG(m,n)

In above formula, P _mV(m, n) and P _bG(m, n) is respectively the pixel value of the image BG of each Moving Objects MV in summarized radio frame and frame background of making a summary; M and n is respectively line number and the row number of pixel in image, and m=1,2 ..., M, n=1,2 ..., N; w ₁, w ₂for weight coefficient, and w ₁+ w ₂=1;

${\mathrm{Corr}}_{(BG,MV)}=\frac{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}{\sqrt{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]}^2\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}^2}}$

In formula, Corr is the coefficient correlation of MV and BG two width source images, for the pixel average of source images MV, for the pixel average of source images BG;

$\left\{\begin{array}{c}{\mathrm{\ω}}_1=\frac{1}{2}(1-\left|Corr\right|)\\ {\mathrm{\ω}}_2=1-{\mathrm{\ω}}_1\end{array}\right.$

$G(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{{(x-x_c)}^2+{(y-y_c)}^2}{2{\mathrm{\σ}}^2}}$

Wherein, σ is the standard deviation of normal distribution, (x _c, y _c) for Gaussian function distribution barycenter.

Embodiment 2:

The track of Moving Objects be according to early stage moving target detection identification and track and extract obtain, wherein, moving object detection cognitive phase is the target detection identification done after Gaussian Background rebuilds extraction motion target area, a grader will identify a kind of object, and then be the track combination optimization by the simulated annealing improved, make the whole track energy of all objects minimum, finally will obtain pedestrian's summary that people and Che separate and vehicle is made a summary.

Traditional monitoring summarized radio, when carrying out the choosing of summarized radio background, selects a Gauss model set up original video often.But in the monitor video scene of reality, the background of original video is change, such as, have a static vehicle at current video frame, resting is known from experience by as background, along with the movement of time shaft, this static vehicle there occurs motion, and it is no longer just background.If choose be initial stationary vehicle time background, the background of this original video namely extracted, when make an abstract frame background, cannot reflect real video monitoring scene like this.Based on addressing this problem, propose a kind of strategy that can dynamically update the background of summary frame herein, the background making each summary frame can be the background meeting the movement locus of object in original video.

When object trajectory and background are directly pasted, the background of following the tracks of object agglomerate and the summary frame obtained is not mated, and has and significantly splices vestige.Based on addressing this problem, devising a kind of Trace Formation strategy based on Gaussian Profile herein, reaching the seamless spliced effect of Moving Objects and frame of video background.

As shown in Figure 3 in the algorithm of target detection schematic diagram of rank adaptive model, the HOG feature that have selected cart headstock and dolly vehicle body image is as Expressive Features and carry out the detection identification of vehicle vehicle in conjunction with the full features training sorting technique of Linear SVM, simultaneously, target vehicle region was extracted roughly by GMM (GaussianMixtureModel) before HOG feature extraction, thus establish rank adaptive model, reduce operand, improve precision.

Target tracking algorism flow chart as shown in Figure 4, the method to be applicable in multiple-camera video camera between two and to carry out the situation of multiple target tracking.

Sdi video target profile as shown in Figure 5, transverse axis represents the initial time of this target, and the longitudinal axis represents each target.Very dense at summarized radio zone line, and all can there is a sparse region in beginning and end, obviously do not meet and be uniformly distributed, the present invention adopts the Kd-Trees method based on cluster, balance the moving target number of each frame in summarized radio, avoid target too concentrated in a certain frame, to such an extent as to be difficult to target not to be fused to overlappingly in summarized radio.Adopt the method, can also fall and the moving target of starting point in adjacent area may be gathered a class, be allowed to condition in summarized radio and in chronological sequence sequentially occur in video streaming.Like the scene dispatched a car in a track, preceding vehicle is first left, after vehicle can not have when overlaid pixel with preceding vehicle and dispatch a car again.So just can avoid not the conflicting in summarized radio of moving target, overlapping or intersect, effectively reducing the probability of conflict.

Sdi video goal programming distribution map as shown in Figure 6, transverse axis represents the initial time of this target, and the longitudinal axis represents each target.With Fig. 5 unlike, when total line number is certain, there is multiple target every provisional capital.Can ensure that the number of targets that each frame occurs is consistent like this, ensure that moving target being uniformly distributed in each frame, make full use of the space of each frame, effectively shorten the length of summarized radio.

Table 1

Conventional method as shown in table 1 and the contrast of to improve one's methods herein, when track number, primary power are identical with termination energy condition, the simulated annealing method convergence rate improved herein is higher than traditional simulated annealing.This is owing to first having carried out clustering planning analysis herein, and the initial distribution making track is enough even, has then carried out again that energy based on simulated annealing is minimum to be solved.Initial distribution enough evenly greatly speeded the convergence rate of simulated annealing.

Do not lose problem for the fast browsing of monitor video and most of action message, in conjunction with the feature of monitor video itself, propose a kind of summarized radio generation technique framework based on object trajectory, and achieve the key technology in this technological frame.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (10)

1. a generation system for summarized radio, is characterized in that: comprise and detect recognition and tracking extraction module (1), Combinatorial Optimization module (2), frame of video background module (3), track background stickup module (4), summarized radio module (5);

Described detection recognition and tracking extraction module (1), for detecting the Moving Objects identified in motion target area, follows the tracks of and movement locus extraction detecting the Moving Objects identified;

Described Combinatorial Optimization module (2), for carrying out Combinatorial Optimization to the movement locus extracted;

Described frame of video background module (3), for dynamically updating the video background in motion target area;

Described track background stickup module (4), for according to the movement locus after Combinatorial Optimization by Moving Objects with upgrade after video background carry out mate pasting and merge, generation summarized radio frame;

Described summarized radio module (5), for generating summarized radio according to summarized radio frame.

2. the generation system of a kind of summarized radio according to claim 1, is characterized in that: described detection recognition and tracking extraction module (1) identifies the Moving Objects in motion target area by setting up rank adaptive model; Described rank adaptive model is using the HOG feature of cart headstock and dolly vehicle body image as Expressive Features, and the detection identification of vehicle vehicle is carried out in conjunction with the full features training sorting technique of Linear SVM, meanwhile, before HOG feature extraction, target vehicle region is extracted roughly by GMM.

3. the generation system of a kind of summarized radio according to claim 2, it is characterized in that: described SVM full features training sorting technique is specially: can assign to say for linear two classes, find hyperplane, i.e. a linear classifier, all training samples correctly can be classified, namely meet:

$y_i\[w^T{x}_i+b\]>0,\∀i=1,...,n$

Wherein

Yi represents the class label giving input data, and xi represents i-th input HOG characteristic vector;

Due to the nonuniqueness of hyperplane, solve optimal hyperlane by the method for quadratic programming;

$\left\{\begin{array}{c}{w}^{*}=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i{y}_i{x}_i\\ b^{*}=y_i-w^{*T}{x}_i\end{array}\right.$

Wherein, α _irepresent fixing Lagrange multiplier, w ^*represent the normal vector of optimal hyperlane, b ^*represent the side-play amount of optimal hyperlane.

4. the generation system of a kind of summarized radio according to claim 1, it is characterized in that: described detection recognition and tracking extraction module (1) carries out moving object tracking by multiple-camera method for tracking target, and described multiple-camera method for tracking target is specially:

$\left\{\begin{array}{c}{p}_{x_1}^{camA/t}(x,y)=H_{camB-camA}\×Foo{t}_A^{camB/t}(x,y)\\ p_{x_2}^{camA/t}(x,y)=H_{camB-camA}\×Foo{t}_B^{camB/t}(x,y)\end{array}\right.$

$\left\{\begin{array}{c}{p}_{x_1}^{camA/t}(x,y)\∈Area\left(x\right)\\ p_{x_2}^{camA/t}(x,y)\∈Area\left(x\right)\end{array}\right.$

Wherein represent the coordinate position of minimum point coordinate in camB of t target A, H _camB-camAthe affine projection matrix of camB to camA scene, represent the coordinate of any point in t camA, Area (x) is ground regional extent shared by target.

5. the generation system of a kind of summarized radio according to claim 1, it is characterized in that: the video background of described frame of video background module (3) dynamically updates and is specially: at each Moving Objects track of a frame of video, video background is voted, select the background of all Moving Objects tracks under original video environment.

6. a kind of generation system of summarized radio according to any one of claim 1 to 5, is characterized in that: the video background after upgrading and movement locus agglomerate background are carried out the pixel fusion algorithm process of Gaussian Profile by described track background stickup module (4).

7. the generation system of a kind of summarized radio according to claim 6, is characterized in that: the image pixel blending algorithm of described Gaussian Profile is specially:

F(m,n)＝ω ₁P _MV(m,n)+ω ₂P _BG(m,n)

In above formula, P _mV(m, n) and P _bG(m, n) is respectively the pixel value of the image BG of each Moving Objects MV in summarized radio frame and frame background of making a summary; M and n is respectively line number and the row number of pixel in image, and m=1,2 ..., M, n=1,2 ..., N; w ₁, w ₂for weight coefficient, and w ₁+ w ₂=1;

${\mathrm{Corr}}_{(BG,MV)}=\frac{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}{\sqrt{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]}^2\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}^2}}$

In formula, Corr is the coefficient correlation of MV and BG two width source images, the pixel average of source images MV, for the pixel average of source images BG;

$\left\{\begin{array}{c}{\mathrm{\ω}}_1=\frac{1}{2}(1-\left|Corr\right|)\\ {\mathrm{\ω}}_2=1-{\mathrm{\ω}}_1\end{array}\right.$

$G(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{{(x-x_c)}^2+{(y-y_c)}^2}{2{\mathrm{\σ}}^2}}$

Wherein, σ is the standard deviation of normal distribution, (x _c, y _c) for Gaussian function distribution barycenter.

8. a generation method for summarized radio, is characterized in that, comprise the following steps:

Step S1. detects the Moving Objects identified in motion target area, follows the tracks of and movement locus extraction detecting the Moving Objects identified;

Step S2. carries out Combinatorial Optimization to the movement locus extracted; Video background in motion target area is dynamically updated;

Step S3. according to the movement locus after Combinatorial Optimization by Moving Objects with upgrade after video background carry out mate pasting and merge, generation summarized radio frame;

Step S4. generates summarized radio according to summarized radio frame.

9. the generation method of a kind of summarized radio according to claim 8, it is characterized in that, detecting the specific implementation of the Moving Objects identified in motion target area in described step S1: by setting up rank adaptive model, the Moving Objects in motion target area is identified; Described rank adaptive model is using the HOG feature of cart headstock and dolly vehicle body image as Expressive Features, and the detection identification of vehicle vehicle is carried out in conjunction with the full features training sorting technique of Linear SVM, meanwhile, before HOG feature extraction, target vehicle region is extracted roughly by GMM.

10. the generation method of a kind of summarized radio according to claim 8 or claim 9, is characterized in that, the specific implementation of described step S3: the pixel fusion algorithm process of the video background after upgrading and movement locus agglomerate background being carried out Gaussian Profile; The image pixel blending algorithm of described Gaussian Profile is specially:

F(m,n)＝ω ₁P _MV(m,n)+ω ₂P _BG(m,n)

In above formula, P _mV(m, n) and P _bG(m, n) is respectively the pixel value of the image BG of each Moving Objects MV in summarized radio frame and frame background of making a summary; M and n is respectively line number and the row number of pixel in image, and m=1,2 ..., M, n=1,2 ..., N; w ₁, w ₂for weight coefficient, and w ₁+ w ₂=1;

${\mathrm{Corr}}_{(BG,MV)}=\frac{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}{\sqrt{\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{MV}(m,n)-\stackrel{\‾}{P_{MV}}\]}^2\underset{m=1}{\overset{M}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\[P_{BG}(m,n)-\stackrel{\‾}{P_{BG}}\]}^2}}$

In formula, Corr is the coefficient correlation of MV and BG two width source images, for the pixel average of source images MV, for the pixel average of source images BG;

$\left\{\begin{array}{c}{\mathrm{\ω}}_1=\frac{1}{2}(1-\left|Corr\right|)\\ {\mathrm{\ω}}_2=1-{\mathrm{\ω}}_1\end{array}\right.$

$G(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{{(x-x_c)}^2+{(y-y_c)}^2}{2{\mathrm{\σ}}^2}}$

Wherein, σ is the standard deviation of normal distribution, (x _c, y _c) for Gaussian function distribution barycenter.