CN107016693B - Particle filter target tracking algorithm combined with effective anomaly point detection - Google Patents
Particle filter target tracking algorithm combined with effective anomaly point detection Download PDFInfo
- Publication number
- CN107016693B CN107016693B CN201710032389.1A CN201710032389A CN107016693B CN 107016693 B CN107016693 B CN 107016693B CN 201710032389 A CN201710032389 A CN 201710032389A CN 107016693 B CN107016693 B CN 107016693B
- Authority
- CN
- China
- Prior art keywords
- tracking
- target
- calculating
- particle
- feature
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a particle filter target tracking algorithm combined with effective anomaly point detection, which is based on a particle filter framework and comprises the steps of firstly decomposing a target template by using PCA (principal component analysis), establishing mutually orthogonal feature spaces, then projecting particles into the feature spaces in subsequent tracking, calculating reconstruction errors of the particles, and estimating the position of a target through probability weighted sum of the particles. And for the estimated tracking result, detecting abnormal points in the tracking result by using a Lorentz estimator, and judging whether the tracking result reaches an updated set threshold value or not by counting the number of the abnormal points. When a certain number of frames are collected, the frames are projected into the original feature space again, and the reconstruction error is calculated. And carrying out secondary PCA decomposition on the reconstruction error, and updating the eigenvector with the largest eigenvalue in the eigenvector into the original eigenspace. The invention can accurately detect whether the updating is needed or not, and update in time, thereby not only avoiding unnecessary updating, but also improving the tracking accuracy.
Description
Technical Field
The invention belongs to a target tracking technology in the field of computer vision, and particularly relates to a particle filter target tracking algorithm combined with effective outlier detection.
Background
With the rapid development of modern technologies, the application of video monitoring systems has been integrated into parking lots and streets of many residential districts, especially in banks, airport security checks and other special occasions related to the safety of people's lives and properties. Therefore, the research of the video monitoring technology becomes a hot spot of the domestic and foreign research. One of the technologies that has received much attention is target tracking. Although many tracking algorithms have been studied at present, the target tracking technology still faces many challenges, for example, when the appearance of a target changes due to illumination, view angle, occlusion, etc., it is a difficult problem of target research to update the appearance model in time and accurately and efficiently estimate the position of the target.
A commonly used target tracking framework comprises two modules, one being a target state estimation module and one being a target appearance model module. The target state estimation mainly estimates the position of a target of a next frame according to a current frame, common algorithms are an optical flow method, mean-shift, particle filtering and the like, and the particle filtering is widely applied due to the characteristics of nonlinearity, non-Gaussian and nonparametric.
In the establishment of the target appearance model, a Principal Component Analysis (PCA) is one of data Analysis tools, and is often used to analyze the correlation between the dimensions of the gray scale features, and the dimensions are reduced by a decorrelation technique to establish the target appearance model. In the framework of particle filtering, PCA firstly models a feature space of a target template, then projects particles to the feature space for reconstruction, judges the similarity between the particles and a target by calculating the reconstruction error of the particles, and finally estimates the position of the target by the probability weighting of the particles. The feature space established by PCA is simple in calculation and high in speed in the tracking process and is not influenced by illumination and noise. But is sensitive to changes in the appearance of the target and therefore the PCA feature space needs to be updated during the tracking process. At present, most PCA-based algorithms have updated modules, but the updating frequency is too high, so that the tracking efficiency is reduced. For example, in the ilv (incremental Learning for visual) algorithm proposed in j.lim 2004 and in the osp (online object tracking with sparse protocols) in Wang D in 2013, PCA is used for appearance modeling, and the previous feature space is updated using incremental PCA, and in both algorithms, the update frequency is directly performed at an update frequency of once every 5 frames, and such update frequency not only causes unnecessary updates, but also greatly increases the complexity of the calculation. If the update frequency is reduced, the update is not timely enough, and the tracking result is wrong. Therefore, how to achieve the optimal update frequency, avoid increasing unnecessary complexity, and avoid the error tracking result caused by the update being not timely enough is a problem that needs to be solved urgently.
Disclosure of Invention
The invention aims at the technical problem that when the feature space established by PCA is updated in target tracking, frequent updating can cause unnecessary calculation, but if the updating frequency is reduced, target tracking loss can be caused due to untimely updating.
In order to solve the problems, the invention provides a particle filter target tracking algorithm combined with effective abnormal point detection, wherein before updating, a judgment criterion is added, abnormal point detection is carried out on a tracking result, the number of abnormal pixels is counted, and whether the abnormal number reaches an updating requirement or not is judged by setting a threshold value through experiments. And once the number of the abnormal points exceeds a threshold value, updating the feature space.
The technical scheme of the invention specifically comprises the following steps:
(1) in the framework of particle filtering, modeling a feature space of a target template by using a PCA method;
(2) in the subsequent tracking, the weight of the particles is calculated by using the feature space, the position of the target is estimated by calculating the probability weighted sum of the particles, and the PCA of the first stage is completed;
(3) using a Lorentz estimator to detect abnormal values of the tracking result of the PCA in the first stage, calculating the number of the abnormal values, and judging whether the number of the abnormal values reaches the updating standard or not by setting a threshold value;
(4) and carrying out secondary PCA decomposition on the tracking result reaching the updating standard.
Further, the step 1 specifically includes the following steps:
(2) Calculating X covariance matrix R ═ XTX;
(3) Calculating an eigenvalue and an eigenvector of the covariance matrix R;
(4) and taking the feature vector with large corresponding feature value as the feature vector space of the target template.
The step 2 comprises the following steps:
(1) when tracking a target, particles are projected into a target feature space, the particles are represented using vectors of the feature space, and a particle reconstruction error is calculatedWherein the content of the first and second substances,y represents a particle, U is the established feature space, c is the coefficient to be calculated;
(2) calculating the coefficient c using least squares estimation, i.e. c-UTy;
(3) And finally, estimating the position of the target in the current frame by calculating the probability weighted sum of the particles.
The step 3 comprises the following steps:
(1) the Lorentz estimator used isBy calculating an objective functionCalculating a coefficient c by y-cU, wherein y is a tracking result, and U is a feature space of the target template;
(2) the minimization problem is solved using a gradient descent method, in the course of which the coefficient c is in accordance withIteration is carried out, the coefficient c is according toIteration is carried out, and when the two parameters are converged, the iteration process is ended;
(3) after iteration is finished, calculating and tracking target reconstruction error by using the coefficient, wherein the error is greater thanThe pixel points are regarded as abnormal points, and the abnormal points smaller than the threshold value can be regarded as normal points;
(4) and finally, calculating all the number of the abnormal points, setting a threshold value, wherein the value can be obtained through experiments, and if the number of the abnormal points is greater than the threshold value, the tracking result needs to be updated.
The performing of the secondary PCA decomposition in the step 4 specifically includes:
(1) collecting the tracking results to be updated, and projecting the tracking results into the feature space again when a certain number of frames are reached;
(2) calculating a reconstruction error of the tracking result by using a least square method, and carrying out secondary PCA decomposition on the error;
(3) and in the feature vectors obtained by decomposition, adding the feature vector with the maximum feature value into the previous feature space to finish one-time updating.
Compared with the prior art, the invention has the beneficial effects that:
the method comprises the steps of 1, using a Lorentz estimator to detect abnormal values of tracking results in the target tracking algorithm, and judging whether updating is needed or not by counting the number of abnormal points. Firstly, the judgment condition is added, on one hand, the updating frequency can be greatly reduced, and on the other hand, whether the updating is needed or not can be accurately detected.
The method for judging the update by calculating the abnormal points is more accurate than the method for directly calculating the square sum of the reconstruction errors of all the pixels, because one part of all the pixels comes from the background, if the background changes, the whole reconstruction error sum is increased, so that unnecessary update is increased, and the Lorentz estimator only concerns the abnormal values, so that the influence of the background on the update is reduced to a certain extent.
Drawings
Fig. 1 is a schematic representation of the geometry of the projection of a particle in feature space U established by PCA.
Fig. 2 is a diagram of a lorentz estimator function and its influence function.
Fig. 3 is a flow chart of the whole algorithm of the present invention.
Fig. 4 is a target feature space established using PCA.
Fig. 5 is a schematic diagram illustrating detection of abnormal points when the target appearance is not changed much.
Fig. 6 is a schematic diagram illustrating detection of abnormal points when the change in the appearance of the target is large.
FIG. 7 is a schematic diagram illustrating detection of an abnormal point when a target is occluded.
FIG. 8 is a graph showing a statistical curve of outliers in each frame of the sequence.
Detailed Description
The following description of the embodiments of the present invention will be made in detail with reference to the accompanying drawings.
The invention provides a particle filter target tracking algorithm combined with effective anomaly point detection, which mainly comprises the following 4 steps:
1. in the framework of particle filtering, the PCA technique is used to model the feature space of the target template.
Firstly, reading in a first frame of a video sequence, manually framing out a target, and performing simple color histogram modeling on the target; then, around the target, particle distribution is set according to the Gaussian distribution with the mean value of 0 and the variance of 5, the similarity between the particles and the target is judged by calculating the Euclidean distance between each particle and the target color histogram, and corresponding weight is given; and finally, calculating the probability weighted sum of the particles to estimate the particles of the next frame. After the tracking result is stored for 5 continuous tracking frames each time, the tracking result is decomposed into 5 mutually perpendicular feature vectors by using PCA, the 5 feature vectors jointly form a feature space U of the target template, and the space perpendicular to the U is expressed as
2. In the subsequent tracking, the weight of the particles is calculated by using the feature space, and the position of the target is estimated by calculating the probability weighted sum of the particles, which is PCA in the first stage;
in subsequent tracking, the particle is projected into this feature space and then the reconstruction error of the particle is calculated. As shown in fig. 1, vectorRepresenting a particle, coefficient c being a vectorThe calculation formula is that c is equal to UTy; the projection of the particle in space U is therefore UUTy, the reconstruction error isFrom reconstruction errorsThe euclidean distance between the particle and the feature space U, that is, the similarity between the particle and the target, can be determined. Then pass throughThe weight of the particles was obtained, where σ was 0.25.
3. Using a Lorentz estimator to detect abnormal points of the tracking result of the PCA in the first stage, calculating the number of the abnormal points, and judging whether the number of the abnormal values reaches the updating standard or not by setting a threshold value;
the tracking result in the step 2 is represented by y, and the anomaly point in y is calculated by using a Lorentz estimator, wherein the calculation formula is as follows:
this parameter is estimated using a gradient descent method, the iterative formula being:
where c is the coefficient of the feature vector, σ refers to the boundary between the anomaly and normal point decisions, and both parameters are to converge in the iteration. The parameters are initialized as: c. C0=[1,1,1,1,1]T,σ0The parameter superscript indicates the number of iterations, 0 being the value of the parameter at which the iteration has not yet started. The step size step of gradient descent is 0.01; if the step length is too large, the gradient descent method is easy to vibrate, if the step length is too small, the descent speed is too slow, and more iteration times are needed for convergence, the step length design in the experiment is obtained through the experiment, and the convergence can be realized only by 15 iterations, in addition, the step length design in the experiment is obtained through the experiment
Besides the parameter c needs to be updated every iteration, the parameter σ needs to be updated every iteration, and the update formula is as follows:
after parameters c and sigma are obtained through an iterative algorithm, a reconstruction error is calculatedWhen in useThe term "normal point" is used to mean abnormal point, and the term "normal point" is used to mean smaller than normal point. Counting the number of the abnormal points, if the number of the abnormal points is larger than 800 (only aiming at the video used at this time, other sequences need to be obtained by experiments), updating the tracking result, and if the number of the abnormal points is smaller than 800, not updating the tracking result.
4. And carrying out secondary PCA decomposition on the tracking result reaching the updating standard.
And collecting the tracking result meeting the updating standard, updating when the tracking result reaches 5 frames, wherein the updating method comprises the steps of projecting the 5 frames into the old feature space by using a least square method again, then calculating a reconstruction error, carrying out secondary PCA decomposition on the reconstruction error, adding the feature vector corresponding to the maximum feature value in the decomposed feature vectors into the original feature space, and finishing updating.
The experimental results are shown in the following figure, fig. 4 is a feature space established for a target template by using PCA, and the feature space is used for subsequent tracking; fig. 5 shows the display of abnormal points detected when the target appearance does not change much, where the number of abnormal points is 462 and σ is 0.194975; fig. 6 shows that the number of abnormal points is 892 and σ is 0.236287 when the change in appearance is large; fig. 7 shows abnormal points when the occlusion occurs, where the number of the abnormal points is 1627, and σ is 0.317551; FIG. 8 is a graphical representation of the statistics of outliers of each frame of the sequence. The two relatively large peaks are when the target is occluded. The occlusion is also regarded as a type of object appearance change, and the occlusion is also updated, so that tracking when the object is occluded can be processed.
It should be noted that the above mentioned is only an embodiment of the present invention, and is not intended to limit the present invention, and the data set and attack mode used in the present embodiment are limited to the embodiment, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (2)
1. A particle filter target tracking algorithm combined with effective outlier detection is characterized by comprising the following steps:
step 1: in the framework of particle filtering, a principal component analysis method is used for modeling a feature space of a target template, and the method specifically comprises the following steps:
(1.2) calculating the covariance matrix R ═ X of XTX;
(1.3) calculating an eigenvalue and an eigenvector of the covariance matrix R;
(1.4) taking the feature vector with large corresponding feature value as a feature vector space of the target template;
step 2: in the subsequent tracking, the weight of the particle is calculated by using the feature space, the position of the target is estimated by calculating the probability weighted sum of the particle, and the principal component analysis of the first stage is completed, which comprises the following steps:
(2.1) when tracking a target, projecting a particle into a target feature space, expressing the particle using a vector of the feature space, and calculating a particle reconstruction errorWherein y represents a particle, U is the established feature space, and c is the coefficient to be calculated;
(2.2) calculating the coefficient c using least squares estimation, i.e. c ═ UTy;
(2.3) finally estimating the position of the target in the current frame by calculating the probability weighted sum of the particles;
and step 3: using a Lorentz estimator to detect abnormal values of the tracking result of the principal component analysis in the first stage, calculating the number of the abnormal values, setting a threshold value to judge whether the number of the abnormal values reaches an updating standard, using y to represent the tracking result in the step 2, using the Lorentz estimator to calculate abnormal points in y, wherein the calculation formula is as follows:
the parameter c is estimated using a gradient descent method, the iterative formula being:
wherein c is the coefficient of the feature vector, σ refers to the boundary of the judgment of the abnormal point and the normal point, both parameters are converged in iteration, and the parameters are initialized as follows: c. C0=[1,1,1,1,1]T,σ0The parameter superscript indicates the number of iterations, 0 is the parameter value when the iteration has not started, and the step size of gradient descent is 0.01; if the step length is too large, the gradient descent method is easy to vibrate, if the step length is too small, the descent speed is too slow, more iteration times are needed for convergence, and in addition, the method has the advantages of high speed, high accuracy and low cost
Besides the parameter c needs to be updated every iteration, the parameter σ needs to be updated every iteration, and the update formula is as follows:
after parameters c and sigma are obtained through an iterative algorithm, a reconstruction error is calculatedWhen in useIf the number of the abnormal points is larger than 800, the tracking result needs to be updated, and if the number of the abnormal points is smaller than 800, the tracking result does not need to be updated;
and 4, step 4: and (4) performing secondary principal component analysis and decomposition on the tracking result reaching the updating standard.
2. The particle filter target tracking algorithm in combination with significant anomaly detection according to claim 1, wherein the performing of the second principal component analysis decomposition in step 4 specifically comprises:
(1) collecting the tracking results to be updated, and projecting the tracking results into the feature space again when a certain number of frames are reached;
(2) calculating a reconstruction error of the tracking result by using a least square method, and performing secondary principal component analysis decomposition on the error;
(3) and in the feature vectors obtained by decomposition, adding the feature vector with the maximum feature value into the previous feature space to finish one-time updating.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710032389.1A CN107016693B (en) | 2017-01-16 | 2017-01-16 | Particle filter target tracking algorithm combined with effective anomaly point detection |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710032389.1A CN107016693B (en) | 2017-01-16 | 2017-01-16 | Particle filter target tracking algorithm combined with effective anomaly point detection |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107016693A CN107016693A (en) | 2017-08-04 |
CN107016693B true CN107016693B (en) | 2020-04-28 |
Family
ID=59440702
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710032389.1A Active CN107016693B (en) | 2017-01-16 | 2017-01-16 | Particle filter target tracking algorithm combined with effective anomaly point detection |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107016693B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112784923A (en) * | 2021-02-07 | 2021-05-11 | 重庆渝抗医药科技有限公司 | Medical abnormal image classification method under big data environment |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102750522A (en) * | 2012-06-18 | 2012-10-24 | 吉林大学 | Method for tracking targets |
CN105139420A (en) * | 2015-08-03 | 2015-12-09 | 山东大学 | Particle filter and perceptual hash-based video target tracking method |
-
2017
- 2017-01-16 CN CN201710032389.1A patent/CN107016693B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102750522A (en) * | 2012-06-18 | 2012-10-24 | 吉林大学 | Method for tracking targets |
CN105139420A (en) * | 2015-08-03 | 2015-12-09 | 山东大学 | Particle filter and perceptual hash-based video target tracking method |
Non-Patent Citations (1)
Title |
---|
A Novel Two-Stage PCA Algorithm for Object Tracking;Yuanyuan Yang等;《International Conference on Communications, Information Management and Network Security》;20160926;第289-292页 * |
Also Published As
Publication number | Publication date |
---|---|
CN107016693A (en) | 2017-08-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Joshi et al. | A survey on moving object detection and tracking in video surveillance system | |
US7376246B2 (en) | Subspace projection based non-rigid object tracking with particle filters | |
CN111311708B (en) | Visual SLAM method based on semantic optical flow and inverse depth filtering | |
CN101924871A (en) | Mean shift-based video target tracking method | |
CN102982537A (en) | Scene change detection method and scene change detection system | |
Karpagavalli et al. | Estimating the density of the people and counting the number of people in a crowd environment for human safety | |
CN111353496B (en) | Real-time detection method for infrared dim targets | |
CN111178261A (en) | Face detection acceleration method based on video coding technology | |
CN111402303A (en) | Target tracking architecture based on KFSTRCF | |
CN107016693B (en) | Particle filter target tracking algorithm combined with effective anomaly point detection | |
CN109472813A (en) | It is a kind of based on background weighting Mean Shift algorithm and Kalman prediction fusion block tracking | |
CN110706254B (en) | Target tracking template self-adaptive updating method | |
Ng et al. | New models for real-time tracking using particle filtering | |
Xu et al. | Improved compressive tracking in surveillance scenes | |
CN108573217B (en) | Compression tracking method combined with local structured information | |
CN107392936B (en) | Target tracking method based on meanshift | |
Hsieh et al. | Grid-based template matching for people counting | |
CN115147613A (en) | Infrared small target detection method based on multidirectional fusion | |
CN110660081B (en) | Target tracking method based on self-adaptive feature selection | |
Huang et al. | Tracking camouflaged objects with weighted region consolidation | |
CN113702941A (en) | Point cloud speed measurement method based on improved ICP (inductively coupled plasma) | |
CN113112506A (en) | Online moving target detection method based on exponential power distribution and matrix decomposition | |
Wang et al. | An improved camshift-based particle filter algorithm for face tracking | |
Wang et al. | Multiple feature fusion for tracking of moving objects in video surveillance | |
CN115965553B (en) | Multi-feature fusion-based face image local enhancement method |
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 |