CN105761279A - Method for tracking object based on track segmenting and splicing - Google Patents

Abstract

The invention discloses a method for tracking an object based on track segmenting and splicing, aimed at addressing the technical problem of poor accuracy of current object tracking methods. The technical solution comprises the following steps: first initializing a track of each object, then conducting effective segmentation on the tracks in accordance with the degree of coincidence between associated results and track history information, performing accurate splicing on the segmented tracks, finally performing splicing on the tracks among framing windows based on the degree of overlapping so as to obtain a final track of the object. The method of the invention addresses the problems of lost tracking and wrong tracking of current multi-object tracking methods in the absence of appearance information of the object, and realizes more accurate multi-object tracking.

Description

Method for tracking target based on track segmentation with splicing

Technical field

The present invention relates to a kind of method for tracking target, particularly relate to a kind of method for tracking target based on track segmentation with splicing.

Background technology

Target following technology all has in military affairs or civil area and is extremely widely applied.Effective data association is the key solving multiple target tracking problem.Data association is deriving from the process that the metric data of single or multiple sensor and known track carry out being mutually paired.The Hungary Algorithm that Edmonds proposes is to solve for the classic algorithm of two dimension assignment problem optimal solution.

Document " Robustobjecttrackingbyhierarchicalassociationofdetection responses; ComputerVision ECCV2008.SpringerBerlinHeidelberg, 2008:788-801 " proposes a kind of method for tracking target based on individual-layer data association.The document, when using Hungary Algorithm to carry out second layer data association, only accounts for the spatial information of target, such as the direction of the distance between target, path segment.When there is target occlusion problem in image sequence, the information of the target that is blocked can produce loss in various degree.If only with spatial information as the input of data association, it is impossible to exactly the position of target is judged, it is easy to produce leakage with or mistake with phenomenon.

Summary of the invention

In order to overcome the deficiency of the accurate rate variance of existing method for tracking target, the present invention provides a kind of method for tracking target based on track segmentation with splicing.First the method initializes the track of each target, then track is effectively split by the matching degree according to association results Yu track historical information, again the track after segmentation is accurately spliced, finally the track between frame window is carried out splicing according to degree of overlapping and obtain the final track of target.The present invention compensate for when lack target appearance information, the leakage of existing multi-object tracking method with and mistake with phenomenon, it is achieved that more accurate multiple target tracking.

The technical solution adopted for the present invention to solve the technical problems is: a kind of method for tracking target based on track segmentation and splicing, is characterized in comprising the following steps:

Step one, one initialization constants f of setting₀, when picture numbers frameNo meets 0 < frameNo < f₀Time, to the object detection results Object in two adjacent two field pictures₁={ obj₁₁,obj₁₂..., obj_1mAnd Object₂={ obj₂₁,obj₂₂..., obj_2m, calculation cost matrix

Wherein, cost_ijFor obj_1iWith obj_2jBetween Euclidean distance, be calculated as follows:

${\mathrm{cost}}_{ij}=\cos t({\mathrm{obj}}_{1i},{\mathrm{obj}}_{2j})=\sqrt{{(x_{1i}-x_{2j})}^2+{(y_{1i}-y_{2j})}^2}---\left(2\right)$

By cost matrixAs the input matrix of data association, with Hungarian Method, obtain coupling matrixIf obj_1iWith obj_2jThe match is successful, then by obj_2jIt is stored in obj_1iTrack in the middle of.By all images in this procedure ergodic initialization constants, obtain the initialization path segment tracklet of each target_i.With method of least square, path segment being carried out conic fitting and try to achieve fitting parameter param, the curvilinear equation of matching is: Ax²+ Bx+C=0.

As picture numbers frameNo > f₀Time, adjacent two two field picture Hungary Algorithms are carried out data association, and judges whether the path segment after association meets segmentation condition.Segmentation condition is:

$\begin{array}{ccc}\&ForAll;{\mathrm{obj}}_{2n},& {\cos }_{ij}<sizeThreshold& (n=1,2,3...n,n\&NotEqual;j)\end{array}---\left(3\right)$

simi(param(tracklet_i),param(tracklet_i+obj_2j))>simithreshold(4)

Wherein, sizethreshold is the minimum dimension of target, and simithreshold is the similarity threshold of path segment.If obj_1iWith obj_2jThe match is successful, if meeting segmentation condition, then splits obj at present frame_1iPath segment, namely terminate tracklet_i, by obj associated with it_2jAs emerging target.By said process, the every two field picture in traversal frame window, namely obtains the path segment set split in present frame window successively.

Step 2, to input path segment set construct its cost matrix cost_n*n.Cost in matrix_ijFor tracklet_iWith tracklet_jBetween distinctiveness ratio, be calculated as follows:

cost_ij=w₁*dist_ij+w₂*frameGap_ij+w₃*direDiff_ij+w₄*veloDiff_ij(5)

Wherein, dist_ij, frameGap_ij, direDiff_ij, veloDiff_ijRespectively tracklet_iWith tracklet_jBetween space length, frame is poor, and speed difference and direction are poor, are respectively calculated as follows:

${\mathrm{dist}}_{ij}=\sqrt{{(x_{startj}-x_{endi})}^2+{(y_{startj}-y_{endi})}^2}---\left(6\right)$

Wherein, (x_startj,y_startj) for tracklet_jTrack initial frame in target location, (x_endi,y_endi) for tracklet_iTrack end frame in target location.

frameGap_ij=startFrame_j-endFrame_i(7)

Wherein, startFrame_jFor tracklet_jTrack initial frame number, endFrame_iFor tracklet_iTrack end frame sequence number.

direDiff_ij=cos (Line (tracklet_i))-cos(Line(tracklet_j))(8)

Wherein, cos (Line (tracklet_i)) for tracklet_iThe cosine value of orbit tangent.

veloDiff_ij=| velo_i-velo_j|(9)

Wherein, velo_iFor tracklet_iIn target velocity, velo_jFor in target velocity.

According to frameGap_ijJudge whether have blank frame between the path segment being successfully associated every a pair.If frameGap_ij=0, then according to frame sequential by this path segment splicing to being successfully associated.If frameGap_ij≠ 0, then need the position calculating target at disappearance frame place, splice according still further to frame sequential.Target is at the position (x at blank frame miss place_miss,y_miss) historical information of application target estimates:

x_miss=(x_estimatei+x_estimatej)/2(10)

x_estimatei=endLocation_i+velo_i*(miss-endFrame_i)(11)

x_estimatej=startLocation_j+velo_j*(startFrame_j-miss)(12)

Wherein, x_estimateiAnd x_estimatejRespectively according to track tracklet_iAnd tracklet_jThe target that goes out of data-evaluation in the position of the x coordinate at miss place.velo_iAnd velo_jRespectively according to tracklet_iAnd tracklet_jTrace information repeatedly estimate the target the tried to achieve average speed at this section of track.Estimate the target position at each blank frame place by this process, finally three sections of tracks are spliced to form long path segment according to frame sequential.

Overlap between step 3, each two consecutive frame window is overlap frame, carries out the screening of preliminary candidate track accordingly:

$iscandidate\{\begin{array}{cc}=0& trackletOverlap<overlapThreshold\\ =1& trackletOverlap\&GreaterEqual;overlapThreshold\end{array}---\left(13\right)$

After obtaining the candidate tracks of every section of track, retraining with other space time informations of track, that tries to achieve every section of track finally splices track.Spliced track is the arithmetic average of target location in two sections of tracks in the target location at overlapping frame overlap place.

The invention has the beneficial effects as follows: first initialize the track of each target, then track is effectively split by the matching degree according to association results Yu track historical information, again the track after segmentation is accurately spliced, finally the track between frame window is carried out splicing according to degree of overlapping and obtain the final track of target.The present invention compensate for when lack target appearance information, the leakage of existing multi-object tracking method with and mistake with phenomenon, it is achieved that more accurate multiple target tracking.

Below in conjunction with detailed description of the invention, the present invention is elaborated.

Detailed description of the invention

The present invention specifically comprises the following steps that based on the method for tracking target of track segmentation with splicing

1. in frame window, path segment generates and segmentation.

Set an initialization constants f₀, when picture numbers frameNo meets 0 < frameNo < f₀Time, to the object detection results Object in two adjacent two field pictures₁={ obj₁₁,obj₁₂..., obj_1mAnd Object₂={ obj₂₁,obj₂₂..., obj_2m, calculation cost matrix

Wherein, cost_ijFor obj_1iWith obj_2jBetween Euclidean distance, be calculated as follows:

${\mathrm{cost}}_{ij}=\cos t({\mathrm{obj}}_{1i},{\mathrm{obj}}_{2j})=\sqrt{{(x_{1i}-x_{2j})}^2+{(y_{1i}-y_{2j})}^2}---\left(2\right)$

By cost matrixAs the input matrix of data association, with Hungarian Method, obtain coupling matrixIf obj_1iWith obj_2jThe match is successful, then by obj_2jIt is stored in obj_1iTrack in the middle of.By all images in this procedure ergodic initialization constants, obtain the initialization path segment tracklet of each target_i.With method of least square, path segment is carried out conic fitting and try to achieve fitting parameter param (tracklet_i), the curvilinear equation of matching is: Ax²+ Bx+C=0.

As picture numbers frameNo > f₀Time, adjacent two two field picture Hungary Algorithms are carried out data association, and judges whether the path segment after association meets segmentation condition.Segmentation condition is:

$\begin{array}{ccc}\&ForAll;{\mathrm{obj}}_{2n},& {\cos }_{ij}<sizeThreshold& (n=1,2,3...n,n\&NotEqual;j)\end{array}---\left(3\right)$

simi(param(tracklet_i),param(tracklet_i+obj_2j))>simithreshold(4)

Wherein, sizethreshold is the minimum dimension of target, and simithreshold is the similarity threshold of path segment.If obj_1iWith obj_2jThe match is successful, if meeting segmentation condition, then splits obj at present frame_1iPath segment, namely terminate tracklet_i, by obj associated with it_2jAs emerging target.By said process, the every two field picture in traversal frame window, namely obtains the path segment set split in present frame window successively.

2. path segment splicing in frame window.

First, the path segment set of input is constructed its cost matrixCost in matrix_ijFor tracklet_iWith tracklet_jBetween distinctiveness ratio, be calculated as follows:

cost_ij=w₁*dist_ij+w₂*frameGap_ij+w₃*direDiff_ij+w₄*veloDiff_ij(5)

Wherein dist_ij, frameGap_ij, direDiff_ij, veloDiff_ijRespectively tracklet_iWith tracklet_jBetween space length, frame is poor, and speed difference and direction are poor, are respectively calculated as follows:

${\mathrm{dist}}_{ij}=\sqrt{{(x_{startj}-x_{endi})}^2+{(y_{startj}-y_{endi})}^2}---\left(6\right)$

Wherein, (x_startj,y_startj) for tracklet_jTrack initial frame in target location, (x_endi,y_endi) for tracklet_iTrack end frame in target location.

frameGap_ij=startFrame_j-endFrame_i(7)

Wherein, startFrame_jFor tracklet_jTrack initial frame number, endFrame_iFor tracklet_iTrack end frame sequence number.

direDiff_ij=cos (Line (tracklet_i))-cos(Line(tracklet_j))(8)

Wherein, cos (Line (tracklet_i)) for tracklet_iThe cosine value of orbit tangent.

veloDiff_ij=| velo_i-velo_j|(9)

Wherein, velo_iFor tracklet_iIn target velocity, velo_jFor in target velocity.

According to frameGap_ijJudge whether have blank frame between the path segment being successfully associated every a pair.If frameGap_ij=0, then according to frame sequential by this path segment splicing to being successfully associated.If frameGap_ij≠ 0, then need the position calculating target at disappearance frame place, splice according still further to frame sequential.Target is at the position (x at blank frame miss place_miss,y_miss) historical information of application target estimates:

x_miss=(x_estimatei+x_estimatej)/2(10)

x_estimatei=endLocation_i+velo_i*(miss-endFrame_i)(11)

x_estimatej=startLocation_j+velo_j*(startFrame_j-miss)(12)

Wherein, x_estimateiAnd x_estimatejRespectively according to track tracklet_iAnd tracklet_jThe target that goes out of data-evaluation in the position of the x coordinate at miss place.velo_iAnd velo_jRespectively according to tracklet_iAnd tracklet_jTrace information repeatedly estimate the target the tried to achieve average speed at this section of track.Estimate the target position at each blank frame place by this process, finally three sections of tracks are spliced to form long path segment according to frame sequential.

3. the splicing of long path segment and track output between frame window.

Overlap between each two consecutive frame window is overlap frame, therefore in two frame windows, the track of same target has certain overlap.Carry out the screening of preliminary candidate track accordingly:

$iscandidate\{\begin{array}{cc}=0& trackletOverlap<overlapThreshold\\ =1& trackletOverlap\&GreaterEqual;overlapThreshold\end{array}---\left(13\right)$

After obtaining the candidate tracks of every section of track accordingly, retrain with other space time informations of track, such as veloDiff_ijAnd direDiff_ij, that tries to achieve every section of track finally splices track.Spliced track is the arithmetic average of target location in two sections of tracks in the target location at overlapping frame overlap place.

Claims (1)

1. the method for tracking target based on track segmentation with splicing, it is characterised in that comprise the following steps:

Step one, one initialization constants f of setting₀, when picture numbers frameNo meets 0 < frameNo < f₀Time, to the object detection results Object in two adjacent two field pictures₁={ obj₁₁,obj₁₂..., obj_1mAnd Object₂={ obj₂₁,obj₂₂..., obj_2m, calculation cost matrix

Wherein, cost_ijFor obj_1iWith obj_2jBetween Euclidean distance, be calculated as follows:

${\mathrm{cost}}_{ij}=\cos t({\mathrm{obj}}_{1i},{\mathrm{obj}}_{2j})=\sqrt{{(x_{1i}-x_{2j})}^2+{(y_{1i}-y_{2j})}^2}---\left(2\right)$

By cost matrix cost_m*nAs the input matrix of data association, with Hungarian Method, obtain coupling matrix associate_m*n；If obj_1iWith obj_2jThe match is successful, then by obj_2jIt is stored in obj_1iTrack in the middle of；By all images in this procedure ergodic initialization constants, obtain the initialization path segment tracklet of each target_i；With method of least square, path segment being carried out conic fitting and try to achieve fitting parameter param, the curvilinear equation of matching is: Ax²+ Bx+C=0；

As picture numbers frameNo > f₀Time, adjacent two two field picture Hungary Algorithms are carried out data association, and judges whether the path segment after association meets segmentation condition；Segmentation condition is:

$\&ForAll;{\mathrm{obj}}_{2n},{\mathrm{cost}}_{ij}<sizeThreshold,(n=1,2,3...n,n\&NotEqual;j)---\left(3\right)$

simi(param(tracklet_i),param(tracklet_i+obj_2j))>simithreshold(4)

Wherein, sizethreshold is the minimum dimension of target, and simithreshold is the similarity threshold of path segment；If obj_1iWith obj_2jThe match is successful, if meeting segmentation condition, then splits obj at present frame_1iPath segment, namely terminate tracklet_i, by obj associated with it_2jAs emerging target；By said process, the every two field picture in traversal frame window, namely obtains the path segment set split in present frame window successively；

Step 2, to input path segment set construct its cost matrix cost_n*n；Cost in matrix_ijFor tracklet_iWith tracklet_jBetween distinctiveness ratio, be calculated as follows:

cost_ij=w₁*dist_ij+w₂*frameGap_ij+w₃*direDiff_ij+w₄*veloDiff_ij(5)

Wherein, dist_ij, frameGap_ij, direDiff_ij, veloDiff_ijRespectively tracklet_iWith tracklet_jBetween space length, frame is poor, and speed difference and direction are poor, are respectively calculated as follows:

${\mathrm{dist}}_{ij}=\sqrt{{(x_{startj}-x_{endi})}^2+{(y_{startj}-y_{endi})}^2}---\left(6\right)$

Wherein, (x_startj,y_startj) for tracklet_jTrack initial frame in target location, (x_endi,y_endi) for tracklet_iTrack end frame in target location；

frameGap_ij=startFrame_j-endFrame_i(7)

Wherein, startFrame_jFor tracklet_jTrack initial frame number, endFrame_iFor tracklet_iTrack end frame sequence number；

direDiff_ij=cos (Line (tracklet_i))-cos(Line(tracklet_j))(8)

Wherein, cos (Line (tracklet_i)) for tracklet_iThe cosine value of orbit tangent；

veloDiff_ij=| velo_i-velo_j|(9)

Wherein, velo_iFor tracklet_iIn target velocity, velo_jFor in target velocity；

According to frameGap_ijJudge whether have blank frame between the path segment being successfully associated every a pair；If frameGap_ij=0, then according to frame sequential by this path segment splicing to being successfully associated；If frameGap_ij≠ 0, then need the position calculating target at disappearance frame place, splice according still further to frame sequential；Target is at the position (x at blank frame miss place_miss,y_miss) historical information of application target estimates:

x_miss=(x_estimatei+x_estimatej)/2(10)

x_estimatei=endLocation_i+velo_i*(miss-endFrame_i)(11)

x_estimatej=startLocation_j+velo_j*(startFrame_j-miss)(12)

Wherein, x_estimateiAnd x_estimatejRespectively according to track tracklet_iAnd tracklet_jThe target that goes out of data-evaluation in the position of the x coordinate at miss place；velo_iAnd velo_jRespectively according to tracklet_iAnd tracklet_jTrace information repeatedly estimate the target the tried to achieve average speed at this section of track；Estimate the target position at each blank frame place by this process, finally three sections of tracks are spliced to form long path segment according to frame sequential；

Overlap between step 3, each two consecutive frame window is overlap frame, carries out the screening of preliminary candidate track accordingly:

$iscandidate\left\{\begin{array}{cc}=0& trackletOverlap<overlapThreshold\\ =1& trackletOverlap\&GreaterEqual;overlapThreshold\end{array}\right.---\left(13\right)$

After obtaining the candidate tracks of every section of track, retraining with other space time informations of track, that tries to achieve every section of track finally splices track；Spliced track is the arithmetic average of target location in two sections of tracks in the target location at overlapping frame overlap place.