CN106780567B - Immune particle filter extension target tracking method fusing color histogram and gradient histogram - Google Patents
Immune particle filter extension target tracking method fusing color histogram and gradient histogram Download PDFInfo
- Publication number
- CN106780567B CN106780567B CN201611050101.5A CN201611050101A CN106780567B CN 106780567 B CN106780567 B CN 106780567B CN 201611050101 A CN201611050101 A CN 201611050101A CN 106780567 B CN106780567 B CN 106780567B
- Authority
- CN
- China
- Prior art keywords
- particle
- target
- particles
- color
- time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Abstract
The invention discloses an immune particle filter extended target tracking method fusing color and gradient histograms, which comprises the steps of firstly, manually selecting a target to be tracked, and randomly sampling and establishing initial particles and weights according to the prior distribution of a target state; secondly, randomly sampling in the neighborhood of the target to be tracked by taking the target to be tracked as the center to obtain a new particle set, and respectively calculating the observed values of the color and the gradient direction to obtain a particle weight; then, the number of effective particles is estimated according to the particle weight, if the number of effective particles is smaller than a threshold value, an immune particle set optimization process is started, and if the number of effective particles is larger than the threshold value, a resampling process is started; and finally, estimating the state of the particles, outputting a tracking result, and carrying out the next moment until the video is finished. The method disclosed by the invention integrates the color histogram and the gradient direction histogram, can fully utilize the spatial structure information to express the target characteristics, and introduces an immune optimization algorithm under the condition of sample loss, so that the diversity of particles is effectively increased, and the tracking result is more stable.
Description
Technical Field
The invention relates to an extended target tracking method, which is characterized in that an immune particle filter extended target tracking method fusing color and gradient histograms is used for image processing, computer vision and target detection and positioning. Belonging to the technical field of target detection in a tracking system.
Background
Target tracking is always the research focus in the field of computer vision and is widely applied to safety monitoring, target reconnaissance, aviation guidance and the like. The difficulties of current tracking are mainly reflected in: 1) the target motion information is easy to be confused under the interference of a complex background; 2) when the target is shielded by the obstacle, the tracking loss is easy to occur; 3) for a multi-target scene, the problems of track crossing and the like are easy to occur. Target tracking methods generally include deterministic and probabilistic tracking algorithms, where target tracking by finding the best matching region in an image is called deterministic tracking, such as mean shift; the process of predicting the target state under the bayesian filtering framework is called probabilistic tracking, such as kalman filtering, particle filtering, etc. In the face of a plurality of problems of nonlinearity and non-Gaussian in reality, the particle filter can realize the stable tracking of the target. However, when the target is interfered by backgrounds with similar colors, the target characteristics cannot be completely described only by means of single characteristics, and gradients, textures and the like contain corresponding target characteristics, different characteristics are shown in different environments, the characteristics can be fused into target tracking, and when the target is shielded by an obstacle, particle degradation can be caused, the expression capability of the particles on the target state is seriously influenced, and aiming at the problem of sample poverty, the immune algorithm is added before resampling, so that the diversity of the particles is effectively improved, the target is stably tracked, and the current engineering application requirements are well met.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: aiming at the defects of the prior art, the method for tracking the extended target by the immune particle filter fused with the color histogram and the gradient histogram is provided, the color histogram and the gradient histogram are fully fused, the situation that the environment change cannot be well adapted by only depending on the color characteristic is effectively improved, meanwhile, the diversity of particles is improved by adding the immune particle algorithm, and the stable tracking of the target is realized.
In order to realize the purpose, the invention adopts the technical scheme that: an immune particle filter extended target tracking method fusing color and gradient histograms comprises the following steps:
step one, manually selecting a target to be tracked, and establishing initial particles and weights according to target state prior distribution random sampling, wherein the initial particles and the weights are used as { x }k i,wk iDenotes xk iThe state of the ith particle at time k, wk i1/N is the weight corresponding to the ith particle at the moment k, and N is the number of the particles;
step two, randomly sampling in the neighborhood of the target to be tracked selected in the step one by taking the target to be tracked as the center to obtain a new particle set, and respectively calculating the observed values of the color and the gradient direction to obtain the weight of the particles, wherein the specific steps are as follows:
(21) defining a color observation:
in the formula, λcRepresenting a color control parameter, Dc(pc,qc) Is the Bhattacharyya distance, p, between the color of the candidate region and the color of the target regioncAs candidate region color distribution vector, qcAs a target area color distribution vector, PcThe larger the value is the color observed value, the more similar the candidate area color is to the target template.
(22) Defining gradient direction observations:
in the formula, λTRepresenting a gradient direction control parameter, DT(pT,qT) Is the Bhattacharyya distance, p, between the gradient of the candidate region and the gradient of the target regionTAs candidate region gradient distribution vector, qTAs a gradient distribution vector of the target region, PTIs a gradient direction observation value;
(23) and (3) fusing the color observation value and the gradient direction observation value to calculate the weight value of the particle:
wk i＝wk-1 i[αPc+(1-α)PT]
where α denotes a constant, k denotes the time, i denotes the number of particles (i is 1,2,3, … N), and w denotesk-1 iIs the weight value, w, corresponding to the ith particle at the moment of k-1k iIs the weight value, P, corresponding to the ith particle at the moment kcAs a color observation value, PTIs a gradient direction observation value;
step three, estimating the number of effective particles by adopting the particle weight obtained in the step two, and entering an immune particle set optimization process if the number of effective particles is smaller than a preset threshold, wherein the specific steps are as follows:
(31) estimating the effective particle number:
where k is the time number, i is the number of particles (i ═ 1,2,3, … N), and wk iThe weight value corresponding to the ith particle at the moment k, T represents the size of a preset threshold, N is the number of the particleseIs the effective particle number.
(32) The immune particle set optimization process mainly comprises the following five steps: fitness calculation, memory unit update, antibody concentration regulation, crossover and variation:
(3.2.1) fitness calculation
F(i)＝1/ei
Wherein i is the number of particles (i is 1,2,3, … N), eiThe mean square error of the state estimate from the true value.
And (3.2.2) updating the memory unit, replacing the antibody with low fitness by the antibody with high fitness calculated in (3.2.1), retaining the antibody with high fitness in the memory unit, and retaining the particles with highest affinity in the original antibody during replacement to keep the diversity of the particles.
(3.2.3) antibody concentration regulation, which mainly depends on an antibody concentration regulation mechanism and utilizes a promotion and inhibition mechanism to select antibodies, thereby keeping the diversity of the antibodies and realizing the self-regulation function in individuals. Antibody concentration regulation is specifically expressed as follows:
wherein i is the number of particles (i is 1,2,3, … N), gamma and β are constants respectively, N is the number of particles, p isfiTo adapt the probability, pdiTo suppress the concentration probability, piThe probability of selecting an antibody is determined,
Cirepresenting the similarity ratio between the antibodies, and F (i) is a fitness function.
(3.2.4) antibody crossover, 2 particles (x) were randomly selected from the set of particlesk m,xk n)，xk mM particle at time k, xk nAnd performing cross processing for the nth particle at the time k according to the following formula:
where α is a constant, k is the time, i is the number of particles (i is 1,2,3, … N), and xk mM particle at time k, xk nη for the nth particle at time k obeys a normal distribution N (0, σ) with a mean of 0 and a variance of σ,
the m-th particle at the time k after the cross processing,
the n-th particle at the time k after the crossover processing.
(3.2.5) antibody variation, randomly selecting 1 particle x from the set of particlesk i，xk iFor the ith particle at time k, mutation processing is performed according to the following formula:
where k is the time number, i is the number of particles (i ═ 1,2,3, … N), and xk iη for the ith particle at time k obeys a normal distribution N (0, σ) with a mean of 0 and a variance of σ,
the particle is the ith particle at the k moment after mutation treatment;
step four, if the number of the effective particles in the step three is larger than a preset threshold value, entering a resampling process;
and step five, outputting a tracking result, entering the next moment until the video is finished, performing target state estimation on the particle set obtained in the step three or the step four, and outputting the tracking result, wherein the state estimation adopts the following definition:
(51) state estimation
Where k is the time number, i is the number of particles (i ═ 1,2,3, … N), xk iThe state of the ith particle at time k, wk iIs the weight corresponding to the ith particle at the moment k, N is the number of the particles, XkSet of samples at time k, E (X)k) For a set of samples X at time kkThe corresponding state estimate mean.
Compared with the prior art, the invention has the beneficial effects that:
(1) according to the invention, the immune optimization algorithm is adopted to optimize the particle set, and the particles can be degraded when the target is shielded by the barrier.
(2) The invention integrates the color and gradient histogram characteristics, the color histogram can describe the target globally, and the gradient direction histogram contains corresponding structural information, so that the situation that the environment change can not be well adapted by only depending on the color characteristics is improved.
Drawings
FIG. 1 is a flow chart of a method implementation of the present invention;
fig. 2 is a tracking result obtained by applying a standard particle filter to an actual scene sequence 1 according to the present invention, where fig. 2(a) is a tracking result obtained by applying a standard particle filter to a 236 th frame of the actual scene sequence 1, fig. 2(b) is a tracking result obtained by applying a standard particle filter to a 280 th frame of the actual scene sequence 1, fig. 2(c) is a tracking result obtained by applying a standard particle filter to a 281 th frame of the actual scene sequence 1, and fig. 2(d) is a tracking result obtained by applying a standard particle filter to a 289 th frame of the actual scene sequence 1;
fig. 3 is a tracking result of applying an immune particle filter to an actual scene sequence 1 according to the present invention, where fig. 3(a) is a tracking result of applying an immune particle filter to a 236 th frame of the actual scene sequence 1, fig. 3(b) is a tracking result of applying an immune particle filter to a 280 th frame of the actual scene sequence 1, fig. 3(c) is a tracking result of applying an immune particle filter to a 281 th frame of the actual scene sequence 1, and fig. 3(d) is a tracking result of applying an immune particle filter to a 289 th frame of the actual scene sequence 1;
FIG. 4 is a graph of error curves of the present invention for tracking a target by respectively using a standard particle filter and an immune particle filter for an actual scene sequence 1;
fig. 5 is a tracking result of an immune particle filter that depends only on color features for the actual scene sequence 1 according to the present invention, where fig. 5(a) is a tracking result of an immune particle filter that depends only on color features for the 500 th frame of the actual scene sequence 1, fig. 5(b) is a tracking result of an immune particle filter that depends only on color features for the 530 th frame of the actual scene sequence 1, fig. 5(c) is a tracking result of an immune particle filter that depends only on color features for the 570 th frame of the actual scene sequence 1, and fig. 5(d) is a tracking result of an immune particle filter that depends only on color features for the 769 th frame of the actual scene sequence 1;
fig. 6 is a tracking result of an immune particle filter employing a fused color and histogram of gradient directions for an actual scene sequence 1 according to the present invention, where fig. 6(a) is a tracking result of an immune particle filter employing a fused color and histogram of gradient directions for a 500 th frame of the actual scene sequence 1, fig. 6(b) is a tracking result of an immune particle filter employing a fused color and histogram of gradient directions for a 530 th frame of the actual scene sequence 1, fig. 6(c) is a tracking result of an immune particle filter employing a fused color and histogram of gradient directions for a 570 th frame of the actual scene sequence 1, and fig. 6(d) is a tracking result of an immune particle filter employing a fused color and histogram of gradient directions for a 769 th frame of the actual scene sequence 1;
FIG. 7 is an error curve diagram of tracking a target by respectively using an immune particle filter only depending on color features and an immune particle filter fusing color and gradient direction histograms for an actual scene sequence 1 according to the present invention;
fig. 8 is a tracking result of an immune particle filter employing a fused color and gradient direction histogram for an actual scene sequence 2 according to the present invention, where fig. 8(a) is a tracking result of an immune particle filter employing a fused color and gradient direction histogram for a 97 th frame of an actual scene sequence 1, fig. 8(b) is a tracking result of an immune particle filter employing a fused color and gradient direction histogram for a 99 th frame of an actual scene sequence 1, fig. 8(c) is a tracking result of an immune particle filter employing a fused color and gradient direction histogram for a 103 th frame of an actual scene sequence 1, and fig. 8(d) is a tracking result of an immune particle filter employing a fused color and gradient direction histogram for a 108 th frame of an actual scene sequence 1.
Detailed Description
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.
Since the advent of particle filtering, particle filtering has become the mainstream method for solving the problems of parameter estimation and state filtering of nonlinear, non-gaussian dynamic systems, and there are many key problems to be solved in current particle filtering, for example, when a target is disturbed by backgrounds with similar colors, the target features cannot be fully described by only relying on a single feature, and gradients, textures and the like contain corresponding target features, which show different characteristics in different environments, and these features can be fused into target tracking, and when the target is shielded by an obstacle, particle degradation can be caused. Therefore, aiming at the problems, the invention integrates the color and gradient histogram characteristics, well represents the characteristics of the target, and introduces the immune optimization algorithm to effectively solve the problem of sample loss.
The invention relates to an implementation of an immune particle filter extended target tracking method integrating color and gradient histograms, wherein an input image is an extended target image in an actual scene.
As shown in fig. 1, the present invention provides an immune particle filter extended target tracking method with color and gradient histograms fused, which includes the following steps:
step one, manually selecting a target to be tracked, and establishing initial particles and weights according to target state prior distribution random sampling, wherein the initial particles and the weights are used as { x }k i,wk iDenotes xk iThe state of the ith particle at time k, wk i1/N is the weight corresponding to the ith particle at the moment k, and N is the number of the particles;
step two, randomly sampling in the neighborhood of the target to be tracked selected in the step one by taking the target to be tracked as the center to obtain a new particle set, and respectively calculating the observed values of the color and the gradient direction to obtain the weight of the particles, wherein the specific steps are as follows:
(21) defining a color observation:
in the formula, λcRepresenting a color control parameter, Dc(pc,qc) Is the Bhattacharyya distance, p, between the color of the candidate region and the color of the target regioncAs candidate region color distribution vector, qcAs a target area color distribution vector, PcThe larger the value is the color observed value, the more similar the candidate area color is to the target template.
(22) Defining gradient direction observations:
in the formula, λTRepresenting a gradient direction control parameter, DT(pT,qT) Is the Bhattacharyya distance, p, between the gradient of the candidate region and the gradient of the target regionTAs candidate region gradient distribution vector, qTAs a gradient distribution vector of the target region, PTIs a gradient direction observation value;
(23) and (3) fusing the color observation value and the gradient direction observation value to calculate the weight value of the particle:
wk i＝wk-1 i[αPc+(1-α)PT]
where α denotes a constant, k denotes the time, i denotes the number of particles (i is 1,2,3, … N), and w denotesk-1 iIs the weight value, w, corresponding to the ith particle at the moment of k-1k iIs the weight value, P, corresponding to the ith particle at the moment kcAs a color observation value, PTIs a gradient direction observation value;
step three, estimating the number of effective particles by adopting the particle weight obtained in the step two, and entering an immune particle set optimization process if the number of effective particles is smaller than a preset threshold, wherein the specific steps are as follows:
(31) estimating the effective particle number:
where k is the time number, i is the number of particles (i ═ 1,2,3, … N), and wk iThe weight value corresponding to the ith particle at the moment k, T represents the size of a preset threshold, N is the number of the particleseIs the effective particle number.
(32) The immune particle set optimization process mainly comprises the following five steps: fitness calculation, memory unit update, antibody concentration regulation, crossover and variation:
(3.2.1) fitness calculation
F(i)＝1/ei
Wherein i is the number of particles (i is 1,2,3, … N), eiThe mean square error of the state estimate from the true value.
And (3.2.2) updating the memory unit, replacing the antibody with low fitness by the antibody with high fitness calculated in (3.2.1), retaining the antibody with high fitness in the memory unit, and retaining the particles with highest affinity in the original antibody during replacement to keep the diversity of the particles.
(3.2.3) antibody concentration regulation, which mainly depends on an antibody concentration regulation mechanism and utilizes a promotion and inhibition mechanism to select antibodies, thereby keeping the diversity of the antibodies and realizing the self-regulation function in individuals. Antibody concentration regulation is specifically expressed as follows:
wherein i is the number of particles (i is 1,2,3, … N), gamma and β are constants respectively, N is the number of particles, p isfiTo adapt the probability, pdiTo suppress the concentration probability, piThe probability of selecting an antibody is determined,
Cirepresenting the similarity ratio between the antibodies, and F (i) is a fitness function.
(3.2.4) antibody crossover, 2 particles (x) were randomly selected from the set of particlesk m,xk n)，xk mM particle at time k, xk nAnd performing cross processing for the nth particle at the time k according to the following formula:
where α is a constant, k is the time, i is the number of particles (i is 1,2,3, … N), and xk mM particle at time k, xk nη for the nth particle at time k obeys a normal distribution N (0, σ) with a mean of 0 and a variance of σ,
the m-th particle at the time k after the cross processing,
the m-th particle at time k after the crossover processing.
(3.2.5) antibody variation, randomly selecting 1 particle x from the set of particlesk i，xk iFor the ith particle at time k, mutation processing is performed according to the following formula:
where k is the time number, i is the number of particles (i ═ 1,2,3, … N), and xk iη for the ith particle at time k obeys a normal distribution N (0, σ) with a mean of 0 and a variance of σ,
the particle is the ith particle at the k moment after mutation treatment;
step four, if the number of the effective particles in the step three is larger than a preset threshold value, entering a resampling process;
and step five, outputting a tracking result, entering the next moment until the video is finished, performing target state estimation on the particle set obtained in the step three or the step four, and outputting the tracking result, wherein the state estimation adopts the following definition:
(51) state estimation
Where k is the time number, i is the number of particles (i ═ 1,2,3, … N), xk iThe state of the ith particle at time k, wk iFor the ith granule at time kWeight corresponding to the child, N is the number of particles, XkSet of samples at time k, E (X)k) For a set of samples X at time kkThe corresponding state estimate mean.
In order to verify the effectiveness of the invention, 2 real scenes are selected qualitatively to carry out experiments on the invention, wherein fig. 2 is a graph of the tracking result of the scene 1, and as can be seen from the graph, a target can be effectively tracked at first by adopting a standard particle filtering algorithm. But starting from 280 frames, the target is completely lost as it is occluded by a fast-walking person, and this continues until the video ends. The reason for this is that in frames 280 and 281, the target is in a completely shielded state, the acquired particles cannot represent the real state of the target, and in frame 283, the person is far away from the target, and the target tracking is completely lost due to the degradation of the particle set. Fig. 3 is a diagram of a tracking result obtained by applying an immune particle filter to the scene 1, and it is found from fig. 3 that the immune particle filter can still define the target when the 280 th and 281 th frames of targets are completely blocked; fig. 5 shows the tracking result of the immune particle filter only depending on the color features for scene 1, where the airplane starts to descend after 500 th frame and has similar colors to some trees, streets and guardrails, and the immune particle filter algorithm only combining the color features cannot accurately identify and has deviation in tracking due to the similarity of the color features of the two. After frame 530, tracking is lost until the end of the video. The reason is that as long as the target is interfered by a background with similar color, the discrimination is reduced, the background condition is complex, and the target characteristic cannot be effectively represented only by the color characteristic, the tracking precision is not high or the tracking error is caused; fig. 6 is a tracking result diagram of an immune particle filter employing a fused color and gradient direction histogram for scene 1, and it is found after observing fig. 6 that after 500 th frame, the aircraft is influenced by the colors of trees, streets and guardrails in the descending process, and when color tracking is utilized, the information difference between the target and the background can be effectively distinguished by fusing the gradient direction characteristics, and the target tracking can be more accurately realized; FIG. 8 is a graph of the tracking result of the fused color and gradient direction histogram immune particle filter for scene 2, where women are selected as the tracking target, the white frame is the target search frame, the gray frame is the tracking result of the color feature only immune particle filter, and the black frame is the color and gradient direction histogram immune particle filter tracking result, and it is seen from FIG. 8 that when the 97 th frame of women is not blocked, the target is better tracked by the color feature only immune particle filter, the color and gradient direction histogram immune particle filter; in the 99 th frame, the woman is seriously shielded by the barrier, the woman and the barrier can also correctly mark the target position, and the accuracy of the target position is higher; in the 103 th frame, when the colors of the female and the lamp post indicating plate are similar, gray tracking has a point deviation, and black can correctly track the target; in the 108 th frame, after the woman walks through the signboard, the target can be tracked by gray and black, but due to the fact that the gray is interfered by the background in the 103 th frame, tracking errors are introduced, the later-stage precision is slightly deviated, and due to the fact that the black is fused with the color and the gradient direction histogram, the obtaining precision is higher.
In order to quantitatively evaluate the effectiveness of the tracking effect in the invention, the video tracking effect is evaluated by adopting a mean square error, and the specific formula is
Wherein, cx、cyRepresents the tracking target center position of the ith frame, ox、oyRepresenting the true target centre position, err, obtained by handiIndicating the i-th frame tracking offset, erriSmaller indicates more accurate tracking results. Fig. 4 is an error curve diagram of tracking a target by respectively using a standard particle filter and an immune particle filter for the scene 1, where the curve can more clearly reflect the difference between the tracking performances of the two. Before occlusion occurs, the tracking error values of the two algorithms are both smaller than 3.12, but when occlusion occurs, the error of the standard particle filter algorithm is obviously increased, for example, in 280 frames, the tracking error value of the immune particle filter algorithm is equal to 4, the tracking error value of the standard particle filter algorithm is equal to 6, the tracking performance of the standard particle filter algorithm after occlusion is increasingly poor until a target is lost, the current position of the target can be estimated according to the memorized particles when the target is occluded by the immune particle filter algorithm, when the target is occluded, the principle of no sampling is adopted, and after the 283 th frame of people leaves, the algorithm adopts mechanisms of crossing, variation and the like so as to enhance the particle numberThe diversity of the purposes promotes the particles to move towards the direction in which the target can move, realizes the quick positioning of the target and accurately tracks the target; fig. 7 is an error curve diagram of tracking a target by using an immune particle filter only depending on color features and an immune particle filter fusing color and gradient direction histograms respectively for the scene 1, and the result shows: before the frames with similar colors, the tracking error values of the two algorithms are less than 3.2, but when the frames with similar colors appear, the error value of the immune particle filter algorithm only depending on a single color feature is in an increasing trend, for example, in 500 frames, the error value of the immune particle filter algorithm with a single color feature is equal to 6, and the error value of the immune particle filter algorithm fusing color and gradient direction histograms is equal to 3. Therefore, the immune particle filter which only depends on the color characteristics is easy to be interfered by similar background colors to cause large tracking deviation, and the tracking precision of the target can be improved by fusing the color and the gradient direction histogram.
The invention has not been described in detail and is part of the common general knowledge of a person skilled in the art.
It will be appreciated by those skilled in the art that the above embodiments are illustrative only and not intended to be limiting of the invention, and that variations to the above described embodiments may fall within the scope of the appended claims, provided they fall within the true spirit of the invention.
Claims (3)
1. An immune particle filter extension target tracking method fusing color and gradient histograms is characterized in that: the method comprises the following steps:
step one, manually selecting a target to be tracked, and establishing initial particles and weight values according to target state prior distribution random sampling;
step two, randomly sampling in the neighborhood of the target to be tracked selected in the step one by taking the target to be tracked as the center to obtain a new particle set, and respectively calculating the observed values of the color and the gradient direction to obtain the weight of the particles;
in the second step, random sampling is performed in the neighborhood of the target to be tracked, which is selected in the first step, as the center to obtain a new particle set, and the observed values in the color and gradient directions are respectively calculated to obtain the weight of the particles, and the specific steps are as follows:
(21) defining a color observation:
in the formula, λcRepresenting a color control parameter, Dc(pc,qc) Is the Bhattacharyya distance, p, between the color of the candidate region and the color of the target regioncAs candidate region color distribution vector, qcAs a target area color distribution vector, PcThe color observation value is larger, which indicates that the color of the candidate area is more similar to the target template;
(22) defining gradient direction observations:
in the formula, λTRepresenting a gradient direction control parameter, DT(pT,qT) Is the Bhattacharyya distance, p, between the gradient of the candidate region and the gradient of the target regionTAs candidate region gradient distribution vector, qTAs a gradient distribution vector of the target region, PTIs a gradient direction observation value;
(23) and (3) fusing the color observation value and the gradient direction observation value to calculate the weight value of the particle:
wk i＝wk-1 i[αPc+(1-α)PT]
where α represents a constant, k is the time count, i is the particle count, i is 1,2,3, … N, wk-1 iIs the weight value, w, corresponding to the ith particle at the moment of k-1k iIs the weight value, P, corresponding to the ith particle at the moment kcAs a color observation value, PTIs a gradient direction observation value;
step three, estimating the number of effective particles by adopting the particle weight obtained in the step two, and entering an immune particle set optimization process if the number of effective particles is smaller than a preset threshold;
step four, if the number of the effective particles in the step three is larger than a preset threshold value, entering a resampling process;
and step five, outputting a tracking result, entering the next moment until the video is finished, performing target state estimation on the particle set obtained in the step three or the step four, and outputting the tracking result, wherein the state estimation adopts the following definition:
(51) state estimation
Wherein xk iThe state of the ith particle at time k, wk iIs the weight corresponding to the ith particle at the moment k, N is the number of the particles, XkSet of samples at time k, E (X)k) For a set of samples X at time kkThe corresponding state estimate mean.
2. The color and gradient histogram fused immune particle filter extended target tracking method according to claim 1, characterized in that: in the first step, the target to be tracked is manually selected, and initial particles and weight values are established by random sampling according to the prior distribution of the target state, wherein the initial particles and the weight values are { x }k i,wk iDenotes xk iThe state of the ith particle at time k, wk iAnd 1/N is the weight corresponding to the ith particle at the moment k, and N is the number of the particles.
3. The color and gradient histogram fused immune particle filter extended target tracking method according to claim 1, characterized in that: in the third step, the effective particle number is estimated by adopting the particle weight obtained in the second step, and if the effective particle number is smaller than a preset threshold, an immune particle set optimization process is started, and the specific steps are as follows:
(31) estimating the effective particle number:
where k is the time number, i is the particle number, i is 1,2,3, … N, wk iThe weight value corresponding to the ith particle at the moment k, T represents the size of a preset threshold, N is the number of the particleseEffective particle number;
(32) the immune particle set optimization process mainly comprises the following five steps: fitness calculation, memory unit update, antibody concentration regulation, crossover and variation:
(3.2.1) fitness calculation
F(i)＝1/ei
Wherein i is the number of particles, i is 1,2,3, … N, eiThe mean square error of the state estimate and the true value;
(3.2.2) updating the memory unit, namely replacing the antibody with low fitness by the antibody with high fitness calculated in (3.2.1), retaining the antibody with high fitness to the memory unit, and simultaneously retaining the particles with highest affinity in the original antibody during replacement to keep the diversity of the particles;
(3.2.3) antibody concentration regulation, which mainly depends on an antibody concentration regulation mechanism and utilizes a promotion and inhibition mechanism to select antibodies, so as to keep the diversity of the antibodies and realize the self-regulation function in individuals, wherein the antibody concentration regulation is specifically expressed as follows:
wherein i is the number of particles, i is 1,2,3, … N, gamma and β are constants respectively, N is the number of particles, p isfiTo adapt the probability, pdiTo suppress the concentration probability, piIn order to select the probability of an antibody,
Cirepresenting the similarity ratio between antibodies, F (i) is a fitness function;
(3.2.4) antibody crossover, 2 particles (x) were randomly selected from the set of particlesk m,xk n)，xk mM particle at time k, xk nAnd performing cross processing for the nth particle at the time k according to the following formula:
wherein, α1Is a constant, k is the time number, i is the particle number, i is 1,2,3, … N, xk mM particle at time k, xk nη for the nth particle at time k obeys a normal distribution N (0, σ) with a mean of 0 and a variance of σ,
the m-th particle at the time k after the cross processing,
the n-th particle at the k moment after the cross processing;
(3.2.5) antibody variation, randomly selecting 1 particle x from the set of particlesk i，xk iFor the ith particle at time k, mutation processing is performed according to the following formula:
where k is the time number, i is the particle number, i is 1,2,3, … N, xk iη for the ith particle at time k obeys a normal distribution N (0, σ) with a mean of 0 and a variance of σ,
the particle is the ith particle at time k after mutation processing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611050101.5A CN106780567B (en) | 2016-11-24 | 2016-11-24 | Immune particle filter extension target tracking method fusing color histogram and gradient histogram |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611050101.5A CN106780567B (en) | 2016-11-24 | 2016-11-24 | Immune particle filter extension target tracking method fusing color histogram and gradient histogram |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106780567A CN106780567A (en) | 2017-05-31 |
| CN106780567B true CN106780567B (en) | 2020-09-01 |
Family
ID=58912920
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611050101.5A Active CN106780567B (en) | 2016-11-24 | 2016-11-24 | Immune particle filter extension target tracking method fusing color histogram and gradient histogram |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106780567B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107403169B (en) * | 2017-08-08 | 2018-09-28 | 上海识加电子科技有限公司 | Signal lamp detection recognition method and device |
| CN108492314B (en) * | 2018-01-24 | 2020-05-19 | 浙江科技学院 | Vehicle tracking method based on color characteristics and structural features |
| CN110580711A (en) * | 2019-08-23 | 2019-12-17 | 天津大学 | video tracking method adopting particle filtering |
| CN111429488A (en) * | 2020-03-31 | 2020-07-17 | 电子科技大学 | Vehicle continuous tracking method based on traffic video |
| CN112184762A (en) * | 2020-09-05 | 2021-01-05 | 天津城建大学 | Gray wolf optimization particle filter target tracking algorithm based on feature fusion |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831409A (en) * | 2012-08-30 | 2012-12-19 | 苏州大学 | Method and system for automatically tracking moving pedestrian video based on particle filtering |
| CN105389807A (en) * | 2015-10-26 | 2016-03-09 | 南京理工大学 | Particle filter infrared tracking method with fusion of gradient feature and adaptive template |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101872482B (en) * | 2010-06-08 | 2012-05-23 | 东华大学 | Intelligent video tracking method based on immune genetic particle filtering |
| KR101500711B1 (en) * | 2012-01-19 | 2015-03-10 | 한국전자통신연구원 | Method for human tracking using color histograms |
-
2016
- 2016-11-24 CN CN201611050101.5A patent/CN106780567B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831409A (en) * | 2012-08-30 | 2012-12-19 | 苏州大学 | Method and system for automatically tracking moving pedestrian video based on particle filtering |
| CN105389807A (en) * | 2015-10-26 | 2016-03-09 | 南京理工大学 | Particle filter infrared tracking method with fusion of gradient feature and adaptive template |
Non-Patent Citations (2)
| Title |
|---|
| "Head Tracking Using Particle Filter with Intensity Gradient and Color Histogram";Xinyu Xu等;《2005 IEEE International Conference on Multimedia and Expo》;20051024;全文 * |
| "Robust tracking and lost target re-acquisition in video sequences using a combined color-gradient orientations based particle filter and covariance matching";Tarek Benlefki等;《2015 8th International Congress on Image and Signal Processing (CISP)》;20160218;第41-46页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106780567A (en) | 2017-05-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bai et al. | Adaptive dilated network with self-correction supervision for counting | |
| CN106780567B (en) | Immune particle filter extension target tracking method fusing color histogram and gradient histogram | |
| CN103971386B (en) | A kind of foreground detection method under dynamic background scene | |
| WO2020215492A1 (en) | Multi-bernoulli multi-target video detection and tracking method employing yolov3 | |
| CN110660082A (en) | Target tracking method based on graph convolution and trajectory convolution network learning | |
| CN107452015B (en) | Target tracking system with re-detection mechanism | |
| JP4874607B2 (en) | Object positioning device | |
| US20070237359A1 (en) | Method and apparatus for adaptive mean shift tracking | |
| JP2009015827A (en) | Object tracking method, object tracking system and object tracking program | |
| Prokaj et al. | Tracking many vehicles in wide area aerial surveillance | |
| CN105160649A (en) | Multi-target tracking method and system based on kernel function unsupervised clustering | |
| CN110717934A (en) | Anti-occlusion target tracking method based on STRCF | |
| CN106384359B (en) | Motion target tracking method and TV | |
| Sun et al. | Moving vehicle video detection combining ViBe and inter-frame difference | |
| CN114169425A (en) | Training target tracking model and target tracking method and device | |
| Pei et al. | Improved Camshift object tracking algorithm in occluded scenes based on AKAZE and Kalman | |
| Ng et al. | New models for real-time tracking using particle filtering | |
| CN107767401B (en) | Infrared target real-time tracking method and device based on nuclear correlation filtering | |
| Li et al. | Tracking algorithm of multiple pedestrians based on particle filters in video sequences | |
| CN116862832A (en) | Three-dimensional live-action model-based operator positioning method | |
| Castellano et al. | Crowd flow detection from drones with fully convolutional networks and clustering | |
| Ling et al. | Pedestrian detection and feedback application based on YOLOv5s and DeepSORT | |
| Saif et al. | Adaptive long term motion pattern analysis for moving object detection using UAV aerial images | |
| Elassal et al. | Unsupervised crowd counting | |
| CN114399532A (en) | Camera position and posture determining method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |