CN103942536A - Multi-target tracking method of iteration updating track model - Google Patents

Abstract

The invention discloses a multi-target tracking method of an iteration updating track model. The method comprises the steps that firstly, a target detector trained in advance carries out target detection on each image in a video sequence, the track model is initialized through the detection result of the first frame of image, the probability function of a Markov random field is updated through the track model, tracking track fragments are solved through a belief propagation algorithm, the track model is updated again by selecting the confidence tracking track fragment, and the final target track is obtained through iteration. According to the method, through the iteration updating of the track model, the challenge problems of target shielding, missing detection and false detection of the detector, similar targets and the like in the multi-target tracking technology can be solved well, and multi-target reliable and accurate tracking can be achieved.

Description

A kind of iteration is upgraded the multi-object tracking method of locus model

Technical field

The invention belongs to Image processing and compute machine vision field, be specifically related to the multi-object tracking method that a kind of iteration is upgraded locus model.

Background technology

Multiple target tracking refers to a plurality of interesting targets in video sequence, maintains the identity of each target, and the movable information such as position of reasoning target, speed.Multiple target tracking is an important subject of computer vision field, at numerous areas such as intelligent video monitoring, robotics, behavioural analyses, has important using value.

Comparing with monotrack task, there is how challenging difficult point in multiple target tracking.First, the number of target is unknown, and along with the number of target turnover scene objects can change.Secondly, target is often blocked even completely by foreground object or other target partial occlusions, causes algorithm keeps track failure or causes the identity exchange between target.Finally, have the target of similar outward appearance in scene, especially they,, mutually when blocking, are difficult to distinguish each target simultaneously.

Progress along with target detection technique, nearest multi-object tracking method mainly uses the tracking based on detecting, apply an object detector that training in advance is good, on the every width image of video sequence, carry out target detection, track algorithm is mainly that the detection response in time series is carried out to data connection, and the detection response that belongs to same target is connected into a track.As being CN101339608A at publication number, Chinese patent discloses " a kind of method for tracking target and system based on detecting ", this patent is set up tracking target queue, and according to position and yardstick, present frame detection response is mated with object queue, realize target is followed the tracks of and state upgrades, and has strengthened the real-time of following the tracks of.But in target, overlap, when separation or appearance change, reliability is low, follows the tracks of unstable.

Summary of the invention

The object of the invention is to overcome the shortcoming of above-mentioned prior art, the multi-object tracking method that provides a kind of iteration to upgrade locus model, this tracking can high precision, reliably a plurality of targets are followed the tracks of.

For achieving the above object, the multi-object tracking method that iteration of the present invention is upgraded locus model comprises the following steps:

1) by the good object detector of training in advance, the every width image performance objective in video sequence is detected, obtain target detection response, obtain the Feature Descriptor of each target detection response, then video sequence is equally divided into N time window;

2) establish target trajectory model and comprise response target each Feature Descriptor constantly in current time window, according to the target detection of video sequence the first frame, respond to obtain target trajectory model T=(T ₁..., T _k, wherein, T _kbe the locus model of K response target, K responds the quantity of target in the first frame, and the locus model of each response target is corresponding target detection response all, then according to target trajectory model assessment, goes out respectively to respond the position of target in whole time window;

3) each target detection response of establishing in each time window is an observation in markov random file, each is observed and connects a mark, set up markov random file, then by probability function and the level and smooth probability function of target trajectory model modification markov random file, and obtain the marginal probability distribution that the response of each target detection belongs to each locus model, when the marginal probability that belongs to this locus model when target detection response is greater than pre-set threshold value, the mark of this target detection response is revised as to the target sequence number of respective response target;

4) target sequence is number identical and couple together from the target detection response of consecutive frame, forms a target trajectory fragment, then the response target newly obtaining is increased in target trajectory model, and deletes the locus model corresponding to response target of disappearance;

5) repeating step 3 then) and step 4), obtain the target trajectory fragment of all response targets in whole video sequence, then the target trajectory fragment of same target sequence number is coupled together, then after level and smooth and interpolation processing respectively respond the track of target.

Step 1) in, the Feature Descriptor of each target detection response comprises response target's center point coordinate, speed, color histogram and size.

Step 3) concrete steps are: establish the observation y that the target detection response in each time window is markov random file _i, each is observed and connects a mark l _i, the conditional probability that maximizes markov random file is:

$P\left(L\right|Y;T)=\frac{1}{Z}\underset{i}{\mathrm{\&Pi;}}\mathrm{\&phi;}(l_i,y_i)\underset{<i,j>}{\mathrm{\&Pi;}}\mathrm{\&psi;}(l_i,l_j)$

Wherein, Z is normalized factor, the observation set that Y is random field, and the tag set that L is random field, T is locus model, φ (l _i, y _i) be single-point probability function, ψ (l _i, l _j) be level and smooth probability function;

By the single-point probability function φ (l of target trajectory model modification markov random file _i, y _i) and level and smooth probability function ψ (l _i, l _j), recycling faith propagation algorithm obtains each target detection response and belongs to the marginal probability distribution that respectively responds target, when the marginal probability that belongs to this response target when target detection response is greater than pre-set threshold value a, target detection is responded to the target sequence number that corresponding mark is revised as respective response target.

The present invention has following beneficial effect:

Iteration of the present invention is upgraded the multi-object tracking method of locus model when obtaining the track of a plurality of targets, first by the good object detector of training in advance, the every width image performance objective in video sequence is detected, then video sequence is divided into N time window, and obtain target trajectory model according to the target detection response of the first frame in each time window, and then carry out locus model renewal and maximize markov random file conditional probability by iteration, make the more accurate of target trajectory model change, thereby obtain accurate target trajectory, keeping dbjective state to remain unchanged at short notice simultaneously, thereby improve the robustness of target following.

Accompanying drawing explanation

Fig. 1 is schematic flow sheet of the present invention.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further detail:

With reference to figure 1, iteration of the present invention is upgraded the multi-object tracking method of locus model, comprises the following steps:

1) by the good object detector of training in advance, the every width image performance objective in video sequence is detected, obtain target detection response, obtain the Feature Descriptor of each target detection response, then video sequence is equally divided into N time window;

2) establish target trajectory model and comprise response target each Feature Descriptor constantly in current time window, according to the target detection of video sequence the first frame, respond to obtain target trajectory model T={T ₁..., T _k, wherein, T _kbe the locus model of K response target, K responds the quantity of target in the first frame, and the locus model of each response target is corresponding target detection response all, then according to target trajectory model assessment, goes out respectively to respond the position of target in whole time window;

Wherein, locus model T _kcomprise following parameter: represent that target trajectory model prediction current time window is in each residing position of time of day response target, the average velocity that represents response target, the color histogram that represents response target, the size that represents response target.

3) each target detection response of establishing in each time window is an observation in markov random file, each is observed and connects a mark, set up markov random file, then by probability function and the level and smooth probability function of target trajectory model modification markov random file, and obtain the marginal probability distribution that the response of each target detection belongs to each locus model, when the marginal probability that belongs to this locus model when target detection response is greater than pre-set threshold value, the mark of this target detection response is revised as to the target sequence number of respective response target;

4) target sequence is number identical and couple together from the target detection response of consecutive frame, forms a target trajectory fragment, then the response target newly obtaining is increased in target trajectory model, and deletes the locus model corresponding to response target of disappearance;

5) repeating step 3 then) and step 4), obtain the target trajectory fragment of all response targets in whole video sequence, then the target trajectory fragment of same target sequence number is coupled together, then after level and smooth and interpolation processing respectively respond the track of target.

Step 1) in, the Feature Descriptor of each target detection response comprises response target's center point coordinate, speed, color histogram and size.

Step 3) concrete steps are: establish the observation y that the target detection response in each time window is markov random file _i, each is observed and connects a mark l _i, the conditional probability that maximizes markov random file is:

$P\left(L\right|Y;T)=\frac{1}{Z}\underset{i}{\mathrm{\&Pi;}}\mathrm{\&phi;}(l_i,y_i)\underset{<i,j>}{\mathrm{\&Pi;}}\mathrm{\&psi;}(l_i,l_j)$

Wherein, Z is normalized factor, the observation set that Y is random field, and the tag set that L is random field, T is locus model, φ (l _i, y _i) be single-point probability function, ψ (l _i, l _j) be level and smooth probability function;

By the single-point probability function φ (l of target trajectory model modification markov random file _i, y _i) and level and smooth probability function ψ (l _i, l _j), recycling faith propagation algorithm obtains each target detection response and belongs to the marginal probability distribution that respectively responds target, when the marginal probability that belongs to this response target when target detection response is greater than pre-set threshold value a, target detection is responded to the target sequence number that corresponding mark is revised as respective response target.

Wherein, the single-point probability function φ (l of k target _i, y _i), being defined as the similarity that detects response and k object module, similarity is calculated and is divided into center point coordinate, speed, color histogram and size.Level and smooth probability function ψ (l _i, l _j) solve for the adjacent node in neighborhood, frame adjacent node and belong to the adjacent node of same frame before and after comprising.To front and back frame adjacent node, the probability that belongs to same dbjective state when detection response is larger, otherwise less; Adjacent node to same frame, because target can not appear at two places in piece image simultaneously, so detect response, to belong to the probability of same dbjective state less, larger on the contrary.

Claims (3)

1. iteration is upgraded a multi-object tracking method for locus model, it is characterized in that, comprises the following steps:

1) by the good object detector of training in advance, the every width image performance objective in video sequence is detected, obtain target detection response, obtain the Feature Descriptor of each target detection response, then video sequence is equally divided into N time window;

2) establish target trajectory model and comprise response target each Feature Descriptor constantly in current time window, according to the target detection of video sequence the first frame, respond to obtain target trajectory model T={T ₁..., T _k, wherein, T _kbe the locus model of K response target, K responds the quantity of target in the first frame, and the locus model of each response target is corresponding target detection response all, then according to target trajectory model assessment, goes out respectively to respond the position of target in whole time window;

3) each target detection response of establishing in each time window is an observation in markov random file, each is observed and connects a mark, set up markov random file, then by probability function and the level and smooth probability function of target trajectory model modification markov random file, and obtain the marginal probability distribution that the response of each target detection belongs to each locus model, when the marginal probability that belongs to this locus model when target detection response is greater than pre-set threshold value, the mark of this target detection response is revised as to the target sequence number of respective response target;

4) target sequence is number identical and couple together from the target detection response of consecutive frame, forms a target trajectory fragment, then the response target newly obtaining is increased in target trajectory model, and deletes the locus model corresponding to response target of disappearance;

5) repeating step 3 then) and step 4), obtain the target trajectory fragment of all response targets in whole video sequence, then the target trajectory fragment of same target sequence number is coupled together, then after level and smooth and interpolation processing respectively respond the track of target.

2. iteration according to claim 1 is upgraded the multi-object tracking method of locus model, it is characterized in that step 1) in the Feature Descriptor of each target detection response comprise response target's center point coordinate, speed, color histogram and size.

3. iteration according to claim 1 is upgraded the multi-object tracking method of locus model, it is characterized in that step 3) concrete steps be: observe y for one that to establish target detection response in each time window be markov random file _i, each is observed and connects a mark l _i, the conditional probability that maximizes markov random file is:

$P\left(L\right|Y;T)=\frac{1}{Z}\underset{i}{\mathrm{\&Pi;}}\mathrm{\&phi;}(l_i,y_i)\underset{<i,j>}{\&prod;}\mathrm{\&psi;}(l_i,l_j)$

Wherein, Z is normalized factor, the observation set that Y is random field, and the tag set that L is random field, T is locus model, φ (l _i, y _i) be single-point probability function, ψ (l _i, l _j) be level and smooth probability function;

By the single-point probability function φ (l of target trajectory model modification markov random file _i, y _i) and level and smooth probability function ψ (l _i, l _j), recycling faith propagation algorithm obtains each target detection response and belongs to the marginal probability distribution that respectively responds target, when the marginal probability that belongs to this response target when target detection response is greater than pre-set threshold value a, target detection is responded to the target sequence number that corresponding mark is revised as respective response target.