CN108921873B - Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization - Google Patents
Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization Download PDFInfo
- Publication number
- CN108921873B CN108921873B CN201810529460.1A CN201810529460A CN108921873B CN 108921873 B CN108921873 B CN 108921873B CN 201810529460 A CN201810529460 A CN 201810529460A CN 108921873 B CN108921873 B CN 108921873B
- Authority
- CN
- China
- Prior art keywords
- target
- state
- tracking
- lost
- correlation
- 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.)
- Expired - Fee Related
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
- G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Abstract
The invention discloses a Markov decision-making on-line multi-target tracking method based on nuclear correlation filtering optimization, which is characterized in that a detected target is distributed into an activated state, and the target in the activated state is converted into a tracking state or a non-activated state through a pre-trained support vector machine classifier; when the target enters a tracking state, tracking each target on line and constructing an appearance model by combining a TLD tracking algorithm and a kernel correlation filtering tracking algorithm, and judging whether the target keeps tracking continuously or is converted into a lost state by utilizing a high-confidence model updating strategy and median flow tracking stability; and when the target is in a lost state, performing data association by using a similarity function, if the target in the lost state is associated with the detected target, transferring the target in the lost state to a tracking state, otherwise, continuously keeping the target in the lost state, and if the target exceeds the T frames and is in the lost state, transferring the target in the lost state to an inactive state. The invention can make the multi-target tracking performance more robust.
Description
Technical Field
The invention relates to the technical field of computer vision, in particular to a Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization.
Background
The multi-target tracking research of video sequences is an important content in the field of computer vision, and is widely applied to the fields of national defense, video monitoring, intelligent navigation, auxiliary driving, intelligent robots, behavior analysis, video retrieval, biomedicine and the like. The aim of video multi-target tracking is to mark the motion track of each target in a video sequence. However, due to the influence of a plurality of factors such as reduced imaging quality, noise and background interference, changes in the appearance and motion mode of the target, uncertainty of the number of tracked targets, complex and variable occlusion, multi-target tracking algorithm research is a very challenging subject, and a large number of theoretical and technical problems are yet to be solved.
At present, three types of representative methods exist in the field of multi-target tracking. 1) Network flow based tracing. The basic idea of the method is to model the trace as a graph model, wherein each node represents a detection result, each edge represents the transition between two detection results, and the optimal solution is obtained by finding the minimum cost flow in the graph model. 2) Tracking based on energy minimization. The method converts the tracking problem into an energy minimization problem, each solution corresponds to one cost or energy in a solution space, and the algorithm needs to express the cost function and find a proper method to solve the optimal solution. 3) Tracking based on tracking-by detection strategy. The method is based on the premise that the detector can also obtain reliable response under a complex scene, and converts the tracking problem into a two-level data association problem. Firstly, locally associating the detection responses in the continuous frames at a first level to form a local track of the target, and then globally associating the local tracks of the interval frames at a second level to form a complete motion track of the target.
With the continuous improvement of the performance of the detector, the tracking by detection idea is accepted by broad scholars. Since the method relies on the detector, the data correlation algorithm must also consider the influence of detector defects, such as missing detection, false detection, inaccurate detection, etc., on the correlation result. The existing multi-target tracking algorithm is easy to lose under the conditions of shielding, disappearance, reappearance and the like, continuous and effective tracking is not carried out, and the tracking effect is poor.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization, which solves the problem that target tracking is easy to lose under the conditions of shielding, disappearance, reappearance and the like.
In order to achieve the purpose, the technical scheme of the invention is as follows: a Markov decision online multi-target tracking method based on kernel correlation filtering optimization comprises the following steps:
step S1: allocating the detected targets to be in an activated state, and classifying the targets in the activated state into a tracking state or an inactivated state by utilizing a pre-trained support vector machine classifier;
step S2: when the target enters a tracking state, tracking each target on line and constructing an appearance model by combining a TLD tracking algorithm and a kernel correlation filtering tracking algorithm, if the appearance model successfully tracks the target in the next frame, keeping the target in the tracking state, and if not, transferring the state to enter a lost state;
step S3: when the target enters a lost state, determining whether the target in the lost state is associated with a newly detected target by using a similarity function, if so, transferring the target in the lost state to a tracking state, otherwise, continuously keeping the target in the lost state;
step S4: if the target exceeds TlostThe frame is in a lost state, and the target of the lost state is transferred to an inactive state.
Further, the reward function in the activated state is:
wherein phi isActive(s) represents normalized 5-dimensional features, respectively x-coordinate, y-coordinate, width, height, and object detection score of the object,
and bActiveDefining a hyperplane of the support vector machine, wherein a represents action, and when a is equal to a1When y (a) is +1, when a is a2When y (a) ═ 1, a1Indicating the transition of the target from the active state to the tracking state, a2Indicating that the target in the active state is turned to the inactive state.
Further, in the TLD tracking algorithm, uniformly distributed points and Harris corner points of the image are taken from the target area of the previous frame as median flow tracking points.
Further, a high confidence model updating strategy is added into the kernel correlation filtering tracking algorithm, the first confidence index is an output maximum response peak value, the second confidence index is an average peak value correlation energy, and the average peak value correlation energy is:
wherein, Fmax，Fmin，Fw,hRespectively representing the highest response, the lowest response and the response at the (w, h) position, w representing the width of the target and h representing the height of the target;
when APCE and FmaxWhen the values exceed the historical mean value by a certain proportion, the appearance model of the kernel correlation filtering tracking algorithm is updated.
Further, the step S2 specifically includes:
computing a normalized cross-correlation similarity score, NCC, between an online learning stored target template and a TLD tracking algorithm predicted target bounding box location1Calculating a normalized cross-correlation similarity score (NCC) between the target template stored in online learning and the target frame position predicted by the kernel correlation filtering tracking algorithm2：
Wherein, I1Representing the target template, I2Indicating the predicted target bounding box position, I, of the TLD tracking algorithm3Representing the position of a target frame predicted by a kernel correlation filtering tracking algorithm, and representing a dot product operation;
setting the position of the target frame with higher normalized cross correlation similarity score with the target template as the final target output position;
calculating the average value of the overlapping areas between the areas containing the targets in the previous K frames and the corresponding detected results:
wherein, o (t)k,dk) Indicating that the kth frame contains an overlap between the region of the target and the corresponding detection result, K being 1, 2.., K;
the reward function in the tracking state is:
wherein e is0，o0，a0，f0As a threshold value of the decision metric, emedFBMedian representing forward and backward error of optical flow for all sample points, when a ═ a3When y (a) is +1, when a is a4When y (a) ═ 1, a3Indicating that the target in the tracking state is kept in the tracking state, a4The target of the tracking state is converted into the loss state, if various decision metrics are met, the target state is kept in the tracking state, otherwise, the transition state enters the loss state.
Further, the step S3 specifically includes:
if the target exceeds TlostIf the frame is in the lost state, transferring the target of the lost state to the inactive state; if the target in the lost state is associated with the newly detected target, the target is transferred to a tracking state, otherwise, the lost state is continuously maintained;
assuming that t is a lost target and d is a detection result of a target, determining whether the target is associated or not associated with the detection result according to (t, d), adopting a two-classifier to solve the problem of target association, and constructing a function:
f(t,d)＝wTφ(t,d)+b
wherein, (w, b) is a parameter of the function, phi (t, d) represents a feature vector of similarity of the lost target t and the detected target d, and the feature vector comprises a forward error, a backward error, a normalized correlation coefficient, a height ratio, a detection confidence coefficient and a distance between the lost target and the detected target; when f (t, d) is more than or equal to 0, the target t is associated with the target d; when f (t, d) < 0, it indicates no correlation, and the corresponding reward function in the lost state is:
where M denotes a detection number associated as data, and a ═ a6When y (a) ═ 1, the data association between the lost target and the detected target is successful, the target enters the tracking state, and a ═ 15When y (a) is-1, the association of the lost object with the detected object data fails, the object remains in the lost state, a5Indicating that the target in the lost state is to be maintained in the lost state, a6Indicating that the target in the lost state is transferred to the tracking state.
Further, the process of learning the parameters (w, b) includes: learning two classifier parameters (w, b) from labeled trace data using reinforcement learning after defining a mapping between states and actions, feature vector phi (t, d) if a false detection target is associated with one target when markov decision correlation results are compared to real datak) Will be added to the negative sample set, if no object is associated, but an object needs to be associated in the real data, then the feature vector phi (t, d)k) Positive samples are added and the parameters (w, b) are retrained each time the samples are updated until all targets are successfully tracked.
Compared with the prior art, the invention has the beneficial effects that:
(1) the Markov decision multi-target tracking framework is improved by adopting an optical flow tracking point set, and Harris angular points with stronger target expression capability are added on the basis of uniformly distributed points, so that the Markov decision multi-target tracking framework can adapt to the influence of illumination change and scale change.
(2) The method designs the normalized cross correlation similarity fraction, utilizes the target template to respectively analyze the target position predicted by the TLD tracking algorithm and the nuclear correlation filter tracking algorithm, optimizes the final target output position and enables the multi-target tracking performance to be more robust.
(3) And a high confidence coefficient model updating strategy is added, the two confidence coefficient indexes respectively reflect the confidence level of the target and the fluctuation degree of the response graph, and only when the two values are larger than the historical mean value in a certain proportion, the appearance model of the KCF tracking algorithm is updated, so that the situation of tracking drift of the KCF algorithm can be reduced.
Drawings
FIG. 1 is a schematic flow chart of the Markov decision online multi-target tracking method based on the kernel correlation filtering optimization.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
As shown in fig. 1, an online multi-target tracking method for markov decision based on kernel correlation filtering optimization includes the following steps:
step S1: allocating the detected targets to be in an activated state, and classifying the targets in the activated state into a tracking state or an inactivated state by utilizing a pre-trained support vector machine classifier;
step S2: when the target enters a tracking state, tracking each target on line and constructing an appearance model by combining a TLD tracking algorithm and a kernel correlation filtering tracking algorithm, if the appearance model successfully tracks the target in the next frame, keeping the target in the tracking state, and if not, transferring the state to enter a lost state;
step S3: when the target enters a lost state, determining whether the target in the lost state is associated with a newly detected target by using a similarity function, if so, transferring the target in the lost state to a tracking state, otherwise, continuously keeping the target in the lost state;
step S4: if the target exceeds TlostThe frame is in a lost state, and the target of the lost state is transferred to an inactive state.
(1) An active state. Detecting the object by the detectorAssigning an active state, classifying the detected target as a tracking or non-active state, phi, using a Support Vector Machine (SVM) classifierActive(s) represents normalized 5-dimensional features, the 5-dimensional features being the position coordinates, width, height, and object detection score of the object, respectively:
in the formula
And bActiveA hyperplane of the SVM is defined. When a ═ a1When y (a) is +1, when a is a2When y (a) ═ 1, a1Indicating the transition of the target from the active state to the tracking state, a2Indicating that the target in the active state is turned to the inactive state.
(2) The state is tracked. In the tracking state, a decision needs to be made whether to continue to remain in the tracked state or transition to the lost state. In the Markov decision process, no matter when the detection result is transferred to the tracked state, once the detection result is transferred, template features are extracted in the detection result area before the transfer, and the template features are used as initial appearance features of the target. And designing an object tracker which combines the KCF algorithm and the TLD algorithm to effectively track each object on line and construct an appearance model, and if the appearance model can successfully track the object in the next frame, keeping the object in a tracked state.
In the TLD tracking algorithm, uniformly distributed tracking points are taken in a target area of a previous frame, the tracking points are tracked through a median flow, tracking points with front-back direction (FB) errors and Normalized Cross Correlation (NCC) lower than the median are filtered, and a new area can be determined by using the remaining stable points to serve as a new position of a target. However, the tracking points generated uniformly include many smooth points, which cannot effectively represent the target and are easily affected by illumination changes and scale changes. Therefore, the uniform tracking points are easy to be mismatched with similar points in the surrounding background in the tracking process or corresponding matched points can not be found directly.
The Harris corner point is an important local feature of the image, and the point with larger image gray scale change in at least two directions is indicated as the corner point. The angular point features are the first choice of many feature matching algorithms because of the abundant information content, can adapt to illumination changes, and are particularly suitable for processing the problems of shielding and geometric deformation.
In the present embodiment, the uniformly distributed points in the conventional TLD are replaced by uniformly generated tracking points and Harris corner point sets to improve the robustness and accuracy of target tracking. When only Harris corner points are used as tracking points, the number of detected Harris corner points is small when the characteristics of the tracked target are not significant, and the target tracking is easy to drift. Therefore, a set of uniformly distributed points and Harris corner points is used as the median flow tracking points.
The KCF (kernel correlation filtering) tracking algorithm applies a ridge regression method to regress to a two-dimensional gaussian distribution of the target in the feature space, and then finds a response peak in the correlation output in the subsequent tracking sequence to locate the position of the target. The KCF tracking algorithm skillfully applies the kernel function and the fast Fourier transform in the operation, so that the real-time performance of the algorithm is improved. And the cyclic matrix is adopted to simulate sampling, so that intensive sampling can be realized, and the discrimination capability of the model is improved. The KCF tracking algorithm does not perform reliability judgment on tracking results in the tracking process, and cannot judge whether the target tracking fails or not, particularly when the target is shielded, the target is judged to be in a lost state.
In this embodiment, a high confidence model updating strategy is added to the KCF tracking algorithm, the first confidence index is an output maximum response peak value, the second confidence index is an average peak-to-correlation energy (APCE), the two confidence indexes respectively reflect a confidence level of a target and a fluctuation degree of a response map, where APCE is:
in the formula Fmax，Fmin，Fw,hRespectively represent the highest responseLowest response and response at (w, h) position. Only when APCE and FmaxThe KCF tracking algorithm appearance model is updated only when the KCF tracking algorithm appearance model is larger than the historical mean value in a certain proportion, and the situation that the KCF tracking algorithm has tracking drift can be reduced.
In this embodiment, the normalized cross-correlation similarity score, NCC, between the target template stored for online learning and the target bounding box position predicted by the TLD tracking algorithm is calculated1Calculating a normalized cross-correlation similarity score (NCC) between the target template stored in online learning and the target frame position predicted by the kernel correlation filtering tracking algorithm2：
Wherein, I1Representing the target template, I2Indicating the predicted target bounding box position, I, of the TLD tracking algorithm3Representing the position of a target frame predicted by a kernel correlation filtering tracking algorithm, and representing a dot product operation;
setting the position of the target frame with higher normalized cross correlation similarity score with the target template as the final target output position;
the tracked object may get an erroneous result from the detector, and it can be known through intuitive judgment that if a tracked object cannot match with the detection frame in consecutive frames, it is highly likely that the object has received an erroneous detection result in the previous frames. It is necessary to check the history information of the target and calculate the overlap o (t) between the area containing the target in the k-th frame and the corresponding detection resultk,dk) Then, calculating the average value of the overlapping areas between the areas containing the target in the previous k tracking frames and the detection result:
the value is combined with an APCE model updating strategy and a forward and backward error median value to judge whether the target is successfully tracked or not, so that whether the target keeps a tracking state or is converted into a loss state is judged:
e0，o0，a0，f0threshold for each decision metric, emedFBMedian representing forward and backward error of optical flow for all sample points, when a ═ a3When y (a) is +1, when a is a4When y (a) ═ 1, a3Indicating that the target in the tracking state is kept in the tracking state, a4Indicating that the target of the tracking state is transferred to the lost state. If the decision metrics are met, the target state will remain in the tracked state, otherwise the transition state enters the lost state.
When the target is shifted from the lost state to the tracked state, the original tracking template is replaced by the corresponding detection result, so that the accumulation of errors can be avoided.
(3) The state is lost. In the lost state, the markov decision process needs to determine whether the target is still in the lost state or whether the target is transitioned to the tracking state or the inactive state. If the target exceeds TlostIf the frame is in the lost state, transferring the target of the lost state to the inactive state; if the target in the lost state is associated with the newly detected target, the target is transferred to the tracking state, otherwise the lost state is maintained [7 ]]. Assuming that t is a lost target and d is a detection result of one target, it can be determined whether the target is associated (y ═ 1) or not associated (y ═ 1) to the detection result according to (t, d), and a two-classifier is adopted to solve the target association problem, and a function is constructed:
f(t,d)＝wTφ(t,d)+b
where (w, b) is a parameter of the function, and phi (t, d) represents a feature vector of similarity between the tracking target t and the detection target d.
Calculating the similarity between the target template and the detected target uses feature vectors of data association including Forward and Backward (FB) error, Normalized Correlation Coefficient (NCC), height ratio, detection confidence, distance between target and detection. When f (t, d) ≧ 0, it means that target t and target d are associated (y ═ 1); when f (t, d) < 0, it means that there is no correlation (y ═ 1). The corresponding loss reward function is defined as:
where M represents the number of detections associated as data. a ═ a6When y (a) is +1, the data association between the lost target and the detected target is successful, and the target enters a tracking state. a ═ a5When y (a) is-1, the association of the lost object with the detected object data fails, the object remains in the lost state, a5Indicating that the target in the lost state is to be maintained in the lost state, a6Indicating that the target in the lost state is transferred to the tracking state.
In this embodiment, the learning of the strategy is a process of learning parameters (w, b), and after defining the mapping between the states and the actions, the two classifier parameters (w, b) are learned from the labeled trace data by reinforcement learning, and when the MDP correlation result is compared with the real data, if the error detection target is correlated with one target, the feature vector phi (t, d) is obtainedk) Will be added to the negative sample set, if no object is associated, but an object needs to be associated in the real data, then the feature vector phi (t, d)k) Positive samples are added and the parameters (w, b) are retrained each time the samples are updated until all targets are successfully tracked.
The Markov decision multi-target tracking algorithm is easy to lose the target under the conditions of target shielding, target disappearance, target reappearance and the like, and continuous and effective tracking cannot be carried out. By utilizing the strong discrimination capability of the nuclear correlation filter, the invention provides a Markov decision-making on-line multi-target tracking algorithm based on the nuclear correlation filtering optimization. Enhancing the feature expression of a tracked target by combining a kernel correlation filter, firstly improving a tracking point set, and adding Harris angular points with stronger target expression capability on the basis of uniformly distributing points; designing a normalized cross correlation similarity score, and respectively comparing and analyzing the target position predicted by TLD tracking prediction and KCF tracking algorithm by using a target template stored in online learning to optimize the final target output position; and (3) adding a high-confidence-degree model updating strategy into the kernel correlation filter, and judging whether the target keeps tracking continuously or is converted into a lost state by combining the median flow tracking stability as a tracker judgment standard.
The above are preferred embodiments of the present invention, and all changes made according to the technical scheme of the present invention that produce functional effects do not exceed the scope of the technical scheme of the present invention belong to the protection scope of the present invention.
Claims (7)
1. A Markov decision online multi-target tracking method based on kernel correlation filtering optimization is characterized by comprising the following steps:
step S1: allocating the detected targets to be in an activated state, and classifying the targets in the activated state into a tracking state or an inactivated state by utilizing a pre-trained support vector machine classifier;
step S2: when the target enters a tracking state, tracking each target on line and constructing an appearance model by combining a TLD tracking algorithm and a kernel correlation filtering tracking algorithm, if the appearance model successfully tracks the target in the next frame, keeping the target in the tracking state, and if not, transferring the state to enter a lost state;
step S3: when the target enters a lost state, determining whether the target in the lost state is associated with a newly detected target by using a similarity function, if so, transferring the target in the lost state to a tracking state, otherwise, continuously keeping the target in the lost state;
step S4: if the target exceeds TlostThe frame is in a lost state, and the target of the lost state is transferred to an inactive state.
2. The Markov decision-based online multi-goal tracking method of claim 1, wherein the reward function in the activated state is:
wherein phi isActive(s) represents normalized 5-dimensional features, respectively x-coordinate, y-coordinate, width, height, and object detection score of the object,
and bActiveDefining a hyperplane of the support vector machine, wherein a represents action, and when a is equal to a1When y (a) is +1, when a is a2When y (a) ═ 1, a1Indicating the transition of the target from the active state to the tracking state, a2Indicating that the target in the active state is turned to the inactive state.
3. The markov decision online multi-target tracking method according to claim 1, wherein in the TLD tracking algorithm, uniformly distributed points and Harris corner points of an image are taken as median flow tracking points in a target region of a previous frame.
4. The markov decision-based online multi-target tracking method according to claim 1, wherein a high confidence model update strategy is added to the kernel correlation filtering tracking algorithm, a first confidence measure is an output maximum response peak, a second confidence measure is an average peak correlation energy, and the average peak correlation energy is:
wherein, Fmax，Fmin，Fw,hRespectively representing the highest response, the lowest response and (w)H) response at location, w represents width of target, h represents height of target;
when APCE and FmaxWhen the values exceed the historical mean value by a certain proportion, the appearance model of the kernel correlation filtering tracking algorithm is updated.
5. The Markov decision online multi-target tracking method of claim 4, wherein the step S2 specifically comprises:
computing a normalized cross-correlation similarity score, NCC, between an online learning stored target template and a TLD tracking algorithm predicted target bounding box location1Calculating a normalized cross-correlation similarity score (NCC) between the target template stored in online learning and the target frame position predicted by the kernel correlation filtering tracking algorithm2：
Wherein, I1Representing the target template, I2Indicating the predicted target bounding box position, I, of the TLD tracking algorithm3Representing the position of a target frame predicted by a kernel correlation filtering tracking algorithm, and representing a dot product operation;
setting the position of the target frame with higher normalized cross correlation similarity score with the target template as the final target output position;
calculating the average value of the overlapping areas between the areas containing the targets in the previous K frames and the corresponding detected results:
wherein, o (t)k,dk) Indicating that the kth frame contains the region of the object and the corresponding examinationOverlap between measurements, K ═ 1, 2.., K;
the reward function in the tracking state is:
wherein e is0，o0，a0，f0As a threshold value of the decision metric, emedFBMedian representing forward and backward error of optical flow for all sample points, when a ═ a3When y (a) is +1, when a is a4When y (a) ═ 1, a3Indicating that the target in the tracking state is kept in the tracking state, a4The target of the tracking state is converted into the loss state, if various decision metrics are met, the target state is kept in the tracking state, otherwise, the transition state enters the loss state.
6. The Markov decision online multi-target tracking method of claim 1, wherein the step S3 specifically comprises:
if the target exceeds TlostIf the frame is in a lost state, the target of the lost state is transferred to an inactive state; if the target in the lost state is associated with the newly detected target, the target is transferred to a tracking state, otherwise, the lost state is continuously maintained;
assuming that t is a lost target and d is a detection result of a target, determining whether the target is associated or not associated with the detection result according to (t, d), adopting a two-classifier to solve the problem of target association, and constructing a function:
f(t,d)＝wTφ(t,d)+b
wherein, (w, b) is a parameter of the function, phi (t, d) represents a feature vector of similarity of the lost target t and the detected target d, and the feature vector comprises a forward error, a backward error, a normalized correlation coefficient, a height ratio, a detection confidence coefficient and a distance between the lost target and the detected target; when f (t, d) is more than or equal to 0, the target t is associated with the target d; when f (t, d) < 0, it indicates no correlation, and the corresponding reward function in the lost state is:
where M denotes a detection number associated as data, and a ═ a6When y (a) ═ 1, the data association between the lost target and the detected target is successful, the target enters the tracking state, and a ═ 15When y (a) is-1, the association of the lost object with the detected object data fails, the object remains in the lost state, a5Indicating that the target in the lost state is to be maintained in the lost state, a6Indicating that the target in the lost state is transferred to the tracking state.
7. The Markov decision online multi-target tracking method of claim 6, wherein the process of learning the parameters (w, b) comprises: learning two classifier parameters (w, b) from labeled trace data using reinforcement learning after defining a mapping between states and actions, feature vector phi (t, d) if a false detection target is associated with one target when markov decision correlation results are compared to real datak) Will be added to the negative sample set, if no object is associated, but an object needs to be associated in the real data, then the feature vector phi (t, d)k) Positive samples are added and the parameters (w, b) are retrained each time the samples are updated until all targets are successfully tracked.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810529460.1A CN108921873B (en) | 2018-05-29 | 2018-05-29 | Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810529460.1A CN108921873B (en) | 2018-05-29 | 2018-05-29 | Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108921873A CN108921873A (en) | 2018-11-30 |
| CN108921873B true CN108921873B (en) | 2021-08-31 |
Family
ID=64419584
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810529460.1A Expired - Fee Related CN108921873B (en) | 2018-05-29 | 2018-05-29 | Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921873B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109521420B (en) * | 2018-12-20 | 2022-10-04 | 西安电子科技大学 | Multi-target tracking method based on multi-feature matching |
| CN110097569B (en) * | 2019-04-04 | 2020-12-22 | 北京航空航天大学 | Oil tank target detection method based on color Markov chain significance model |
| CN112581496A (en) * | 2019-09-29 | 2021-03-30 | 四川大学 | Multi-target pedestrian trajectory tracking method based on reinforcement learning |
| CN110689559B (en) * | 2019-09-30 | 2022-08-12 | 长安大学 | Visual target tracking method based on dense convolutional network characteristics |
| CN110751096B (en) * | 2019-10-21 | 2022-02-22 | 陕西师范大学 | Multi-target tracking method based on KCF track confidence |
| CN110955239B (en) * | 2019-11-12 | 2021-03-02 | 中国地质大学（武汉） | Unmanned ship multi-target trajectory planning method and system based on inverse reinforcement learning |
| CN111160190B (en) * | 2019-12-21 | 2023-02-14 | 华南理工大学 | Vehicle-mounted pedestrian detection-oriented classification auxiliary kernel correlation filtering tracking method |
| CN111476825B (en) * | 2020-03-10 | 2022-08-26 | 重庆邮电大学 | Anti-occlusion target tracking method based on multi-example learning and kernel correlation filter |
| CN111368938A (en) * | 2020-03-19 | 2020-07-03 | 南京因果人工智能研究院有限公司 | Multi-target vehicle tracking method based on MDP |
| CN112053385B (en) * | 2020-08-28 | 2023-06-02 | 西安电子科技大学 | Remote sensing video shielding target tracking method based on deep reinforcement learning |
| CN113296540B (en) * | 2021-05-20 | 2022-07-12 | 北京航空航天大学 | Hybrid intelligent following and obstacle avoiding method suitable for indoor unmanned aerial vehicle |
| CN114882068B (en) * | 2022-04-24 | 2023-09-01 | 电子科技大学 | Multi-target tracking method, device, computer equipment and storage medium |
| CN115797647B (en) * | 2022-11-14 | 2023-09-08 | 西安电子科技大学广州研究院 | Target stable tracking method under embedded open environment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106650805A (en) * | 2016-12-13 | 2017-05-10 | 中国科学院自动化研究所 | Visual target tracking method and device |
| CN107967692A (en) * | 2017-11-28 | 2018-04-27 | 西安电子科技大学 | A kind of target following optimization method based on tracking study detection |
| WO2018081156A1 (en) * | 2016-10-25 | 2018-05-03 | Vmaxx Inc. | Vision based target tracking using tracklets |
| CN108053427A (en) * | 2017-10-31 | 2018-05-18 | 深圳大学 | A kind of modified multi-object tracking method, system and device based on KCF and Kalman |
-
2018
- 2018-05-29 CN CN201810529460.1A patent/CN108921873B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018081156A1 (en) * | 2016-10-25 | 2018-05-03 | Vmaxx Inc. | Vision based target tracking using tracklets |
| CN106650805A (en) * | 2016-12-13 | 2017-05-10 | 中国科学院自动化研究所 | Visual target tracking method and device |
| CN108053427A (en) * | 2017-10-31 | 2018-05-18 | 深圳大学 | A kind of modified multi-object tracking method, system and device based on KCF and Kalman |
| CN107967692A (en) * | 2017-11-28 | 2018-04-27 | 西安电子科技大学 | A kind of target following optimization method based on tracking study detection |
Non-Patent Citations (3)
| Title |
|---|
| "Multiple Object Tracking with Kernelized Correlation Filters in Urban Mixed Traffic";Yuebin Yang 等;《2017 14th Conference on Computer and Robot Vision (CRV)》;20180208;209-216 * |
| "基于核相关滤波的高速目标跟踪算法研究与系统实现";刘志强;《中国优秀硕士学位论文全文数据库-信息科技辑》;20170315;第2017年卷(第3期);I138-4852 * |
| "基于深度学习的多目标跟踪算法研究";陆平 等;《中兴通讯技术》;20170831;第23卷(第4期);14-19 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108921873A (en) | 2018-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921873B (en) | Markov decision-making online multi-target tracking method based on kernel correlation filtering optimization | |
| CN110751096B (en) | Multi-target tracking method based on KCF track confidence | |
| CN109434251B (en) | Welding seam image tracking method based on particle filtering | |
| CN107564034A (en) | The pedestrian detection and tracking of multiple target in a kind of monitor video | |
| CN110197502B (en) | Multi-target tracking method and system based on identity re-identification | |
| CN106780557A (en) | A kind of motion target tracking method based on optical flow method and crucial point feature | |
| CN113255611B (en) | Twin network target tracking method based on dynamic label distribution and mobile equipment | |
| CN110084830B (en) | Video moving object detection and tracking method | |
| CN110363165B (en) | Multi-target tracking method and device based on TSK fuzzy system and storage medium | |
| CN113763431B (en) | Target tracking method, system, electronic device and storage medium | |
| CN111354022B (en) | Target Tracking Method and System Based on Kernel Correlation Filtering | |
| CN110349188B (en) | Multi-target tracking method, device and storage medium based on TSK fuzzy model | |
| Tsintotas et al. | DOSeqSLAM: Dynamic on-line sequence based loop closure detection algorithm for SLAM | |
| CN111429485B (en) | Cross-modal filtering tracking method based on self-adaptive regularization and high-reliability updating | |
| CN114972735A (en) | Anti-occlusion moving target tracking device and method based on ROI prediction and multi-module learning | |
| CN113537077B (en) | Label multiple Bernoulli video multi-target tracking method based on feature pool optimization | |
| He et al. | Adaptive model update strategy for correlation filter trackers | |
| CN114627339B (en) | Intelligent recognition tracking method and storage medium for cross border personnel in dense jungle area | |
| CN113759928B (en) | Mobile robot high-precision positioning method for complex large-scale indoor scene | |
| Zhu et al. | Visual tracking with dynamic model update and results fusion | |
| Pan et al. | Learning to Track by Bi-Directional Long Short-Term Memory Networks | |
| CN117148309B (en) | Millimeter wave radar human body sensing method and system applied to community grid inspection | |
| Chen et al. | Target Tracking Algorithm Based on Kernel Correlation Filter with Anti-occlusion Mechanisms | |
| CN114782495B (en) | Multi-target tracking method, system and computer storage medium | |
| Wu et al. | Online multi-object tracking based on improved kernel correlation filter |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210831 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |