CN115830075A - Hierarchical association matching method for pedestrian multi-target tracking - Google Patents

Hierarchical association matching method for pedestrian multi-target tracking Download PDF

Info

Publication number
CN115830075A
CN115830075A CN202310132470.2A CN202310132470A CN115830075A CN 115830075 A CN115830075 A CN 115830075A CN 202310132470 A CN202310132470 A CN 202310132470A CN 115830075 A CN115830075 A CN 115830075A
Authority
CN
China
Prior art keywords
detection
target
score
frame
track
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.)
Pending
Application number
CN202310132470.2A
Other languages
Chinese (zh)
Inventor
梅桐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan Guangyinfei Technology Development Co ltd
Original Assignee
Wuhan Guangyinfei Technology Development Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Wuhan Guangyinfei Technology Development Co ltd filed Critical Wuhan Guangyinfei Technology Development Co ltd
Priority to CN202310132470.2A priority Critical patent/CN115830075A/en
Publication of CN115830075A publication Critical patent/CN115830075A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Abstract

The invention discloses a hierarchical association matching method for pedestrian multi-target tracking, which comprises the following steps: the method comprises the following steps: acquiring a target detection frame and a detection score; step two: predicting the position of the track in the current frame; step three: dividing high and low detection frames according to a threshold value; step four: measuring apparent feature distance using cosine distance; step five: measuring a motion feature distance using the DIoU distance; step six: using DIoU distance measurement low score detection and unmatched tracking target; step seven: the invention obtains the boundary frame and the identity ID of the target track, introduces the low-score detection and tracking target matching strategy by adopting a hierarchical matching method so as to recover the shielded pedestrian target in the low-score detection frame, reduces the missing detection between the detection and the track matching and has the characteristic of reducing the missing detection between the detection and the track matching.

Description

Hierarchical association matching method for pedestrian multi-target tracking
Technical Field
The invention relates to the technical field of multi-target tracking of pedestrians, in particular to a hierarchical association matching method for multi-target tracking of pedestrians.
Background
In a monitoring video scene, pedestrian crowding exists at many times, and at the moment, the target is in different states such as local shielding or complete shielding, so that the track information of the tracked target cannot be updated in time.
In order to solve the problem of target interaction occlusion and further improve the accuracy and the real-time performance of a multi-target tracking algorithm, the research direction is focused on establishing more reliable data association measurement. Most methods only keep the detection boxes with detection scores higher than the threshold value to be in data association with the track to obtain the identity information of the track in the current frame. If objects with detection scores below the threshold value (e.g. objects are occluded) are discarded, the non-negligible real object loss and object identity ID switching are brought about. But if all the high and low detection blocks in each frame are considered, more false detections are introduced. Therefore, it is necessary to design a hierarchical association matching method for pedestrian multi-target tracking, which can reduce the missing detection between detection and trajectory matching.
Disclosure of Invention
The invention aims to provide a hierarchical association matching method for pedestrian multi-target tracking so as to solve the problems in the background technology.
In order to solve the technical problems, the invention provides the following technical scheme: a hierarchical association matching method for pedestrian multi-target tracking comprises
The method comprises the following steps: acquiring a target detection frame and a detection score;
step two: predicting the position of the track in the current frame;
step three: dividing high and low detection frames according to a threshold value;
step four: measuring apparent feature distances using cosine distances;
step five: measuring a motion feature distance using the DIoU distance;
step six: using DIoU distance measurement low score detection and unmatched tracking target;
step seven: and obtaining a bounding box and an identity ID of the target track.
According to the above technical solution, the method for obtaining the target detection frame and the detection score further comprises:
first, a video frame is input
Figure SMS_1
Extracting video frame features through a global context information enhancement network of a MOT-CN network
Figure SMS_2
；
Second, the extracted shared features
Figure SMS_3
Decoupling features that are further used for multitasking;
then, the characteristic decoupling unit outputs the specific characteristic
Figure SMS_4
Input to the detection branch,
Figure SMS_5
Inputting the feature into a feature embedding unit;
finally, obtaining a target detection frame according to the center point score, the center point offset and the target boundary frame
Figure SMS_6
And detecting the score
Figure SMS_7
。
According to the above technical solution, the method for predicting the position of the trajectory in the current frame further includes:
using a Kalman filter before separating out the low-resolution detection frame and the high-resolution detection frame
Figure SMS_8
To predict
Figure SMS_9
At the new position of each track in the current frame
Figure SMS_10
The calculation formula is as follows:
Figure SMS_11
=KF_predict(
Figure SMS_12
)
wherein the content of the first and second substances,
Figure SMS_13
is a lost track.
According to the technical scheme, the specific steps of dividing the high and low detection frames according to the threshold value comprise:
according to the set detection score threshold
Figure SMS_15
And
Figure SMS_18
divide all the detection frames into
Figure SMS_21
And
Figure SMS_16
two-part sets, i.e. higher for scores than
Figure SMS_19
Put into the high-resolution detection frame set
Figure SMS_20
(ii) a For the detection of score ranging from
Figure SMS_22
To
Figure SMS_14
The person in (1) is put into the low-score detection frame set
Figure SMS_17
。
According to the above technical solution, the method for measuring the apparent feature distance using cosine distance specifically includes:
firstly, in the high score detection frame of the current frame
Figure SMS_23
And all the tracked target tracks of the previous frame
Figure SMS_24
And a lost track
Figure SMS_25
The data correlation calculation is carried out, namely a high score detection frame is calculated by using cosine distance
Figure SMS_26
And a prediction box of the target trajectory
Figure SMS_27
Degree of appearance similarity of
Figure SMS_28
；
Then, matching between the objects is completed by using the Hungarian algorithm based on the appearance similarity.
According to the above technical solution, the method for measuring a motion feature distance using a DIoU distance further includes:
first, calculate the high score detection box using the DIoU metric
Figure SMS_29
And a prediction box of the remaining target trajectory
Figure SMS_30
Motion feature similarity between them
Figure SMS_31
；
Then, the Hungarian algorithm is used to complete the matching between the objects.
According to the above technical solution, the method for low-score detection and unmatched target tracking by using DIoU distance metric further comprises:
first, the DIoU metric is used to calculate the low score detection box
Figure SMS_32
And a prediction box of the remaining target trajectory
Figure SMS_33
Motion feature similarity between them
Figure SMS_34
；
The unmatched target tracks are then retained in the unmatched tracks
Figure SMS_35
And delete all unmatched low score detection boxes.
According to the technical scheme, the method for obtaining the bounding box and the identity ID of the target track comprises the following steps: for the
Figure SMS_36
When its detection score is higher than a threshold value
Figure SMS_37
If more than two frames exist continuously, a new target track is obtained, and the target track in the current frame is output
Figure SMS_38
And the identity ID of the target.
A hierarchical association matching system using the hierarchical association matching method, characterized in that: the system comprises:
the MOT-CN network architecture is used for obtaining a target detection frame and a detection score;
the track prediction module is used for predicting the position of the track in the current frame;
the dividing module is used for dividing the high-low detection frames according to a threshold value;
the analysis module is used for analyzing the measurement detection target characteristics;
and the matching acquisition module is used for acquiring the bounding box and the identity ID of the target track after matching the detection score.
Compared with the prior art, the invention has the following beneficial effects: according to the invention, the adopted hierarchical matching method introduces a low-score detection and tracking target matching strategy to recover the shielded pedestrian target in the low-score detection frame and reduce the missing detection between detection and track matching.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
fig. 1 is a flowchart of a pedestrian multi-target tracking-oriented hierarchical association matching method according to an embodiment of the present invention;
fig. 2 is a schematic diagram illustrating a module composition of a pedestrian multi-target tracking-oriented hierarchical association matching system according to a second embodiment of the present invention;
fig. 3 is a schematic diagram of the overall structure of the MOT-CN network framework of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
In a first embodiment, fig. 1 is a flowchart of a pedestrian multi-target tracking-oriented hierarchical association matching method provided in an embodiment of the present invention, where this embodiment is applicable to a surveillance video scene, and the method can be executed by the hierarchical association matching system provided in the embodiment of the present invention, as shown in fig. 1, the method specifically includes the following steps:
the method comprises the following steps: and acquiring a target detection frame and a detection score.
In the embodiment of the invention, a target detection frame and a detection score are obtained by using a MOT-CN (Multi-Object Tracking for latency based on CenterNet), wherein the MOT-CN network consists of four parts: the device comprises a feature extraction unit, a feature decoupling unit, a detection branch and feature embedding. MOT-CN with video frame
Figure SMS_39
For input, where H and W respectively represent input video frames
Figure SMS_40
Height and width of (a).
Illustratively, first, a video frame is input
Figure SMS_41
Extracting video frame characteristics through a Global Context Enhancement Network (GCEN) Network of MOT-CN
Figure SMS_42
. The GCEN network comprises ResNet-34, a deep aggregation (DLA) variant, a context-aware enhancement (CEM) module and a channel attention-directed (CAM) module. The CEM module of GCEN aims at enhancing high-level semantic features and transmitting the enhanced semantic information to a lower-level feature level, exploring a large amount of context information from multiple receptive fields; meanwhile, when merging feature information in top-down paths, one CAM bank is added in each cross-connect to optimize the final integrated features at each level to mitigate aliasing effects. The overall feature extraction process of GCEN can be expressed as
Figure SMS_43
Wherein, in the step (A),
Figure SMS_44
representing a shared feature extraction operation.
Second, the extracted shared features
Figure SMS_46
The features that are further used for multitasking are decoupled. The characteristic decoupling unit consists of two branches, each branch comprises a kernel with the size of
Figure SMS_49
The convolution layer is used for decomposing the learned shared characteristics into detection specific information and ReID information so as to solve the problem of multi-task optimization contradiction. As shown by the grey decoupling block (Decouple block) in FIG. 3, the characteristic decoupling input is
Figure SMS_52
The outputs of the two branches being respectively
Figure SMS_47
And
Figure SMS_50
the process can be expressed as:
Figure SMS_53
,
Figure SMS_54
wherein
Figure SMS_45
And
Figure SMS_48
are all shown as
Figure SMS_51
The convolution operation of (1).
Then, the characteristic decoupling unit outputs the specific characteristic
Figure SMS_56
Input to the detection branch,
Figure SMS_60
Input to the feature embedding unit. The detection branch unit comprises three parallel branches which are respectively used for estimating the position of a central point of a target, the offset of the center of the target and the size of a bounding box of the target, so that the related target can be positioned according to specific detection information. The characteristics of the output of the detection branch unit are sequentially recorded as a central point position matrix
Figure SMS_64
. Each branch of the detecting branch unit is provided with one
Figure SMS_58
And a convolution kernel of
Figure SMS_59
The convolution kernel generates the final target feature. This process can be expressed as:
Figure SMS_63
、
Figure SMS_67
、
Figure SMS_55
wherein
Figure SMS_62
The convolution operations of the centroid branch, the centroid offset branch, and the bounding box size branch are represented, respectively. The feature embedding unit extracts features by using each target central point so as to be used for appearance feature matching in a subsequent tracking process and improve the tracking robustness. The input features of the feature embedding unit are
Figure SMS_66
The resulting embedded features are noted
Figure SMS_69
Then, there are:
Figure SMS_57
wherein, in the step (A),
Figure SMS_61
is a feature extraction network consisting of
Figure SMS_65
A convolutional layer and
Figure SMS_68
and (4) rolling and laminating the layers.
Finally, obtaining a target detection frame according to the center point score, the center point offset and the target boundary frame
Figure SMS_76
And detecting the score
Figure SMS_71
. Specifically, firstly, sorting is carried out according to the numerical values of the central point score matrix F1, and the first N points are selected to obtain a detection score set
Figure SMS_79
And center point coordinate set
Figure SMS_77
(ii) a Then according to the central point coordinate set in the central point offset matrix F2
Figure SMS_86
The position of (2) is taken out of the offset of the corresponding position
Figure SMS_74
And from the set of center point coordinates in the target bounding box matrix F3
Figure SMS_81
The position of (2) is taken out of the upper and lower width and height values of the corresponding position
Figure SMS_78
(ii) a Finally, calculating the position set of the target detection frame
Figure SMS_83
，
Figure SMS_70
Where t denotes the few frames of the video throughout. Obtained in this step
Figure SMS_85
According to
Figure SMS_73
And a threshold value
Figure SMS_80
And
Figure SMS_72
further will be
Figure SMS_82
Detection frame division high-resolution detection frame
Figure SMS_75
And low score detection frame
Figure SMS_84
。
Step two: and predicting the position of the track in the current frame.
Illustratively, a Kalman filter is used before separating the low-fraction detection frame and the high-fraction detection frame
Figure SMS_87
To predict
Figure SMS_88
At the new position of each track in the current frame
Figure SMS_89
. The calculation formula is as follows:
Figure SMS_90
=KF_predict(
Figure SMS_91
)
wherein the content of the first and second substances,
Figure SMS_92
is a lost track.
Step three: and dividing high and low detection frames according to a threshold value.
Illustratively, in embodiments of the invention, the score threshold is set according to the detection
Figure SMS_94
And
Figure SMS_97
through experimental analysis, taking
Figure SMS_100
=0.4,
Figure SMS_95
=0.1, the tracking performance can be optimized. Divide all the detection frames into
Figure SMS_98
And
Figure SMS_101
two-part sets, i.e. higher than for scores
Figure SMS_103
Put them into high-resolution detection frame set
Figure SMS_93
(ii) a For the detection of score ranging from
Figure SMS_96
To
Figure SMS_99
Put them into the low score detection box set
Figure SMS_102
。
Step four: appearance feature distance is measured using cosine distance.
Firstly, in the high score detection frame of the current frame
Figure SMS_104
And all the tracked target tracks of the previous frame
Figure SMS_105
And a lost track
Figure SMS_106
The data correlation calculation is carried out, namely a high score detection frame is calculated by using cosine distance
Figure SMS_107
And a prediction box of the target trajectory
Figure SMS_108
Degree of appearance similarity of
Figure SMS_109
. Because the apparent vector dimension of the target is higher, the cosine distance still keeps 'same as 1, orthogonal as 0 and opposite as-1' under the condition of high dimension "The Euclidean distance is affected by the dimension, the range is not fixed, and the meaning is fuzzy. Therefore, the invention selects the cosine distance to calculate the appearance similarity P1 of the two targets, and the calculation formula is as follows:
Figure SMS_110
wherein the content of the first and second substances,
Figure SMS_111
representing the appearance characteristics of the detected object,
Figure SMS_112
and representing the appearance characteristics of the tracking target, wherein the appearance characteristics are obtained by the characteristic embedding unit of the first step.
Then, matching between the objects is completed by using the Hungarian algorithm based on the appearance similarity. The Hungarian algorithm mainly finds the best match between the detection target and the tracking target according to the similarity matrix, and keeps the unmatched detection boxes in the high-score detection boxes
Figure SMS_113
In the method, the unmatched target tracks are kept in the unmatched tracks
Figure SMS_114
In (1).
Step five: the motion feature distance is measured using the DIoU distance.
In the embodiment of the invention, firstly, the DIoU measurement is adopted to calculate the high score detection box
Figure SMS_115
And a prediction box of the remaining target trajectory
Figure SMS_116
Motion feature similarity between them
Figure SMS_117
The formula is as follows:
Figure SMS_118
wherein the content of the first and second substances,
Figure SMS_120
is a detection frame
Figure SMS_124
And a prediction block
Figure SMS_127
The cross-over ratio of (a) to (b);
Figure SMS_122
and
Figure SMS_125
=
Figure SMS_128
respectively show the detection frame
Figure SMS_130
And a prediction box
Figure SMS_119
The coordinates of the center point of (a),
Figure SMS_123
is the euclidean distance, the center point of the target in this step is the low dimensional data with dimension 2, the euclidean distance is suitable for low dimensional data measurement, while the appearance vector of step four is the high dimensional data, suitable for cosine distance measurement,
Figure SMS_126
is calculated by the formula
Figure SMS_129
For calculating the distance between the center points;
Figure SMS_121
is the diagonal length of the smallest rectangular box covering both boxes.
Then, the Hungarian algorithm is used to complete the matching between the objects. Will notThe matched detection frame is reserved in the high score detection frame
Figure SMS_131
In the method, the unmatched target tracks are kept in the unmatched tracks
Figure SMS_132
In (1).
Step six: and (4) detecting and unmatching the tracking target by using a DIoU distance measure.
In the embodiment of the invention, the DIoU measurement is firstly adopted to calculate the low score detection box
Figure SMS_133
And a prediction box of the remaining target trajectory
Figure SMS_134
Motion feature similarity between them
Figure SMS_135
The formula is as follows:
Figure SMS_136
wherein the content of the first and second substances,
Figure SMS_139
is a detection frame
Figure SMS_142
And a prediction box
Figure SMS_143
The size of the cross-over ratio of (c),
Figure SMS_138
and
Figure SMS_141
respectively show the detection frame
Figure SMS_144
And a prediction box
Figure SMS_145
The coordinates of the center point of (a),
Figure SMS_137
is the euclidean distance, for calculating the distance between the center points,
Figure SMS_140
is the diagonal length of the smallest rectangular box covering both boxes.
The unmatched target tracks are then retained in the unmatched tracks
Figure SMS_146
And delete all unmatched low score detection boxes.
Step seven: and obtaining a bounding box and an identity ID of the target track.
In the embodiment of the present invention, for
Figure SMS_147
If its detection score is higher than a threshold value
Figure SMS_148
If more than two frames exist continuously, a new target track is obtained, and the target track in the current frame is output
Figure SMS_149
And the identity ID of the target.
The basic idea of the embodiment of the invention is that each detection frame is subjected to detection according to the detection scores
Figure SMS_150
And high score detection threshold
Figure SMS_151
And low score detection threshold
Figure SMS_152
Is divided if
Figure SMS_153
Divide the detection box into high scoreDetect the frame if
Figure SMS_154
The detection frames are divided into low-level detection frames, the high-level detection frames and the low-level detection frames are sequentially subjected to hierarchical association matching, similarity measurement and data association matching is firstly carried out on the high-level detection frames and tracks through appearance features and motion features, similarity calculation is carried out on the remaining unmatched tracks and the low-level detection frames again, and the remaining unmatched tracks and the low-level detection frames are associated, so that pedestrian targets in the low-level detection frames are recovered, and background detection is filtered out. The motion information measures the distance of the motion feature using a DIoU, and simultaneously considers the overlapping area between the target frames and the center point distance of the two bounding boxes when suppressing the redundant boxes.
In a second embodiment, a second embodiment of the present invention provides a hierarchical association matching system for pedestrian multi-target tracking, and fig. 2 is a schematic diagram of modules of the hierarchical association matching system for pedestrian multi-target tracking according to the second embodiment of the present invention, as shown in fig. 2, including:
and the MOT-CN network architecture is used for obtaining a target detection frame and a detection score.
And the track prediction module is used for predicting the position of the track in the current frame.
And the dividing module is used for dividing the high and low detection frames according to the threshold value.
And the analysis module is used for analyzing the measurement detection target characteristics.
And the matching acquisition module is used for acquiring the bounding box and the identity ID of the target track after matching the detection score.
In some embodiments of the invention, the MOT-CN network architecture comprises:
and the characteristic extraction unit is used for extracting the video frame characteristics from the input video frame.
And the characteristic decoupling unit is used for decomposing the learned shared characteristics into specific detection information and ReID information.
And the detection branch unit is used for estimating the position of the center point of the target, the offset of the center of the target and the size of the boundary box of the target.
And the characteristic embedding unit is used for extracting the characteristic by utilizing each target central point.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (9)

1. A hierarchical association matching method for pedestrian multi-target tracking is characterized by comprising the following steps: the method specifically comprises the following steps:
the method comprises the following steps: acquiring a target detection frame and a detection score;
step two: predicting the position of the track in the current frame;
step three: dividing high and low detection frames according to a threshold value;
step four: measuring apparent feature distance using cosine distance;
step five: measuring a motion feature distance using the DIoU distance;
step six: using DIoU distance measurement low score detection and unmatched tracking target;
step seven: and obtaining a bounding box and an identity ID of the target track.
2. The pedestrian multi-target tracking oriented hierarchical association matching method according to claim 1, characterized in that: the method for obtaining the target detection frame and the detection score further comprises the following steps:
first, a video frame is input
Figure QLYQS_1
Extracting video frame characteristics through a global context information enhancement network of a MOT-CN network
Figure QLYQS_2
；
Second, the extracted shared features
Figure QLYQS_3
Decoupling features that are further used for multitasking;
then, the characteristic decoupling unit outputs the specific characteristic
Figure QLYQS_4
Input to the detection branch,
Figure QLYQS_5
Inputting the feature into a feature embedding unit;
finally, obtaining a target detection frame according to the center point score, the center point offset and the target boundary frame
Figure QLYQS_6
And detecting the score
Figure QLYQS_7
。
3. The pedestrian multi-target tracking-oriented hierarchical association matching method according to claim 1, characterized in that: the method for predicting the position of the track in the current frame further comprises the following steps:
using a Kalman filter before separating out the low-resolution detection frame and the high-resolution detection frame
Figure QLYQS_8
To predict
Figure QLYQS_9
At the new position of each track in the current frame
Figure QLYQS_10
The calculation formula is as follows:
Figure QLYQS_11
=KF_predict(
Figure QLYQS_12
)
wherein the content of the first and second substances,
Figure QLYQS_13
is a lost track.
4. The pedestrian multi-target tracking-oriented hierarchical association matching method according to claim 1, characterized in that: the specific steps of dividing the high and low detection frames according to the threshold value comprise:
according to the set detection score threshold
Figure QLYQS_15
And
Figure QLYQS_18
divide all the detection frames into
Figure QLYQS_21
And
Figure QLYQS_16
two-part sets, i.e. higher than for scores
Figure QLYQS_17
Put into the high-resolution detection frame set
Figure QLYQS_20
(ii) a For the detection of score ranging from
Figure QLYQS_22
To
Figure QLYQS_14
The person in (1) is put into the low-score detection frame set
Figure QLYQS_19
。
5. The pedestrian multi-target tracking-oriented hierarchical association matching method according to claim 1, characterized in that: the method for measuring the apparent feature distance by using the cosine distance specifically comprises the following steps:
firstly, in the high score detection frame of the current frame
Figure QLYQS_23
And all the tracked target tracks of the previous frame
Figure QLYQS_24
And a lost track
Figure QLYQS_25
The data correlation calculation is carried out, namely a high score detection frame is calculated by using cosine distance
Figure QLYQS_26
And a prediction box of the target trajectory
Figure QLYQS_27
Degree of appearance similarity of
Figure QLYQS_28
；
Then, matching between the objects is completed by using the Hungarian algorithm based on the appearance similarity.
6. The pedestrian multi-target tracking-oriented hierarchical association matching method according to claim 1, characterized in that: the method of measuring a motion feature distance using a DIoU distance further includes:
first, calculate the high score detection box using the DIoU metric
Figure QLYQS_29
And a prediction box of the remaining target trajectory
Figure QLYQS_30
Motion feature similarity between them
Figure QLYQS_31
；
And then, matching among the targets is completed by using a Hungarian algorithm.
7. The pedestrian multi-target tracking-oriented hierarchical association matching method according to claim 1, characterized in that: the method for detecting and unmatching the tracking target with the DIoU distance metric further comprises the following steps:
the low score detection box is first calculated using the DIoU metric
Figure QLYQS_32
And a prediction box of the remaining target trajectory
Figure QLYQS_33
Motion feature similarity between them
Figure QLYQS_34
；
The unmatched target tracks are then retained in the unmatched tracks
Figure QLYQS_35
And delete all unmatched low score detection boxes.
8. According to claim 1The hierarchical association matching method for pedestrian multi-target tracking is characterized by comprising the following steps: the method for obtaining the bounding box and the identity ID of the target track comprises the following steps: for the
Figure QLYQS_36
When its detection score is higher than a threshold value
Figure QLYQS_37
If more than two frames exist continuously, a new target track is obtained, and the target track in the current frame is output
Figure QLYQS_38
And the identity ID of the target.
9. A hierarchical relevance matching system using the hierarchical relevance matching method of any of the preceding claims 1-8, characterized by: the system comprises:
the MOT-CN network architecture is used for obtaining a target detection frame and a detection score;
the track prediction module is used for predicting the position of the track in the current frame;
the dividing module is used for dividing the high-low detection frames according to the threshold value;
the analysis module is used for analyzing the measurement detection target characteristics;
and the matching acquisition module is used for acquiring the bounding box and the identity ID of the target track after matching the detection score.
CN202310132470.2A 2023-02-20 2023-02-20 Hierarchical association matching method for pedestrian multi-target tracking Pending CN115830075A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310132470.2A CN115830075A (en) 2023-02-20 2023-02-20 Hierarchical association matching method for pedestrian multi-target tracking

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310132470.2A CN115830075A (en) 2023-02-20 2023-02-20 Hierarchical association matching method for pedestrian multi-target tracking

Publications (1)

Publication Number Publication Date
CN115830075A true CN115830075A (en) 2023-03-21

Family

ID=85521810

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310132470.2A Pending CN115830075A (en) 2023-02-20 2023-02-20 Hierarchical association matching method for pedestrian multi-target tracking

Country Status (1)

Country Link
CN (1) CN115830075A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115994928A (en) * 2023-03-24 2023-04-21 知行汽车科技(苏州)股份有限公司 Target tracking method, device, equipment and medium
CN116128932A (en) * 2023-04-18 2023-05-16 无锡学院 Multi-target tracking method
CN116681729A (en) * 2023-06-14 2023-09-01 石家庄铁道大学 Anti-occlusion multi-target tracking method based on track prediction
CN116703983A (en) * 2023-06-14 2023-09-05 石家庄铁道大学 Combined shielding target detection and target tracking method
CN117173221A (en) * 2023-09-19 2023-12-05 浙江大学 Multi-target tracking method based on authenticity grading and occlusion recovery
CN117496446A (en) * 2023-12-29 2024-02-02 沈阳二一三电子科技有限公司 People flow statistics method based on target detection and cascade matching

Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814889A (en) * 2020-07-14 2020-10-23 大连理工大学人工智能大连研究院 Single-stage target detection method using anchor-frame-free module and enhanced classifier
CN112288773A (en) * 2020-10-19 2021-01-29 慧视江山科技(北京)有限公司 Multi-scale human body tracking method and device based on Soft-NMS
CN113129345A (en) * 2021-04-19 2021-07-16 重庆邮电大学 Target tracking method based on multi-feature map fusion and multi-scale expansion convolution
CN113674328A (en) * 2021-07-14 2021-11-19 南京邮电大学 Multi-target vehicle tracking method
CN113743260A (en) * 2021-08-23 2021-12-03 北京航空航天大学 Pedestrian tracking method under dense pedestrian flow condition of subway platform
CN114639032A (en) * 2020-12-15 2022-06-17 上海交通大学 Vehicle detection tracking method, device and equipment for quasi-real-time digital twin display
CN114638855A (en) * 2022-01-21 2022-06-17 山东汇创信息技术有限公司 Multi-target tracking method, equipment and medium
CN114724059A (en) * 2022-03-14 2022-07-08 南京邮电大学 Chain type multi-target tracking method for secondary association low-resolution detection frame
CN114820702A (en) * 2022-04-14 2022-07-29 大连理工大学 Jepsort unmanned aerial vehicle view angle pedestrian multi-target tracking method based on yolov5
CN114972410A (en) * 2022-06-16 2022-08-30 上海影谱科技有限公司 Multi-level matching video racing car tracking method and system
CN115063454A (en) * 2022-08-16 2022-09-16 浙江所托瑞安科技集团有限公司 Multi-target tracking matching method, device, terminal and storage medium
CN115240130A (en) * 2022-06-15 2022-10-25 首都师范大学 Pedestrian multi-target tracking method and device and computer readable storage medium
CN115631214A (en) * 2022-09-29 2023-01-20 深圳市中钞信达金融科技有限公司 Multi-target tracking method and system based on motion information and semantic information
CN115641454A (en) * 2022-10-19 2023-01-24 智道网联科技(北京)有限公司 Target tracking method and device, electronic equipment and computer readable storage medium

Patent Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814889A (en) * 2020-07-14 2020-10-23 大连理工大学人工智能大连研究院 Single-stage target detection method using anchor-frame-free module and enhanced classifier
CN112288773A (en) * 2020-10-19 2021-01-29 慧视江山科技(北京)有限公司 Multi-scale human body tracking method and device based on Soft-NMS
CN114639032A (en) * 2020-12-15 2022-06-17 上海交通大学 Vehicle detection tracking method, device and equipment for quasi-real-time digital twin display
CN113129345A (en) * 2021-04-19 2021-07-16 重庆邮电大学 Target tracking method based on multi-feature map fusion and multi-scale expansion convolution
CN113674328A (en) * 2021-07-14 2021-11-19 南京邮电大学 Multi-target vehicle tracking method
CN113743260A (en) * 2021-08-23 2021-12-03 北京航空航天大学 Pedestrian tracking method under dense pedestrian flow condition of subway platform
CN114638855A (en) * 2022-01-21 2022-06-17 山东汇创信息技术有限公司 Multi-target tracking method, equipment and medium
CN114724059A (en) * 2022-03-14 2022-07-08 南京邮电大学 Chain type multi-target tracking method for secondary association low-resolution detection frame
CN114820702A (en) * 2022-04-14 2022-07-29 大连理工大学 Jepsort unmanned aerial vehicle view angle pedestrian multi-target tracking method based on yolov5
CN115240130A (en) * 2022-06-15 2022-10-25 首都师范大学 Pedestrian multi-target tracking method and device and computer readable storage medium
CN114972410A (en) * 2022-06-16 2022-08-30 上海影谱科技有限公司 Multi-level matching video racing car tracking method and system
CN115063454A (en) * 2022-08-16 2022-09-16 浙江所托瑞安科技集团有限公司 Multi-target tracking matching method, device, terminal and storage medium
CN115631214A (en) * 2022-09-29 2023-01-20 深圳市中钞信达金融科技有限公司 Multi-target tracking method and system based on motion information and semantic information
CN115641454A (en) * 2022-10-19 2023-01-24 智道网联科技(北京)有限公司 Target tracking method and device, electronic equipment and computer readable storage medium

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
YIFU ZHANG 等: ""FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking"", pages 1 - 19 *
YIFU ZHANG 等: "ByteTrack: Multi-Object Tracking by Associating Every Detection Box", pages 1 - 14 *
明悦: ""多源视觉信息感知与识别"", 北京：北京邮电大学出版社, pages: 115 - 116 *

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115994928A (en) * 2023-03-24 2023-04-21 知行汽车科技(苏州)股份有限公司 Target tracking method, device, equipment and medium
CN116128932A (en) * 2023-04-18 2023-05-16 无锡学院 Multi-target tracking method
CN116681729A (en) * 2023-06-14 2023-09-01 石家庄铁道大学 Anti-occlusion multi-target tracking method based on track prediction
CN116703983A (en) * 2023-06-14 2023-09-05 石家庄铁道大学 Combined shielding target detection and target tracking method
CN116703983B (en) * 2023-06-14 2023-12-19 石家庄铁道大学 Combined shielding target detection and target tracking method
CN116681729B (en) * 2023-06-14 2024-01-26 石家庄铁道大学 Anti-occlusion multi-target tracking method based on track prediction
CN117173221A (en) * 2023-09-19 2023-12-05 浙江大学 Multi-target tracking method based on authenticity grading and occlusion recovery
CN117173221B (en) * 2023-09-19 2024-04-19 浙江大学 Multi-target tracking method based on authenticity grading and occlusion recovery
CN117496446A (en) * 2023-12-29 2024-02-02 沈阳二一三电子科技有限公司 People flow statistics method based on target detection and cascade matching
CN117496446B (en) * 2023-12-29 2024-03-15 沈阳二一三电子科技有限公司 People flow statistics method based on target detection and cascade matching

Similar Documents

Publication Publication Date Title
CN115830075A (en) Hierarchical association matching method for pedestrian multi-target tracking
JP6759474B2 (en) Vessel automatic tracking methods and systems based on deep learning networks and average shifts
Xie et al. Real-time illegal parking detection system based on deep learning
CN108304808A (en) A kind of monitor video method for checking object based on space time information Yu depth network
CN109145836B (en) Ship target video detection method based on deep learning network and Kalman filtering
CN106885574A (en) A kind of monocular vision robot synchronous superposition method based on weight tracking strategy
CN110263712A (en) A kind of coarse-fine pedestrian detection method based on region candidate
CN111767847A (en) Pedestrian multi-target tracking method integrating target detection and association
CN113628245B (en) Multi-target tracking method, device, electronic equipment and storage medium
CN113763427B (en) Multi-target tracking method based on coarse-to-fine shielding processing
Han et al. A method based on multi-convolution layers joint and generative adversarial networks for vehicle detection
CN111931571B (en) Video character target tracking method based on online enhanced detection and electronic equipment
Li et al. Bi-directional dense traffic counting based on spatio-temporal counting feature and counting-LSTM network
CN114926859A (en) Pedestrian multi-target tracking method in dense scene combined with head tracking
CN115457082A (en) Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement
Dagvasumberel et al. Railroad near-miss occurrence detection and risk estimation system with data from camera using deep learning
CN112613472B (en) Pedestrian detection method and system based on deep search matching
CN112164097B (en) Ship video detection sample collection method
CN106407989A (en) Fast automatic density clustering-based detection method for scale-variable infrared small target
Tian et al. Pedestrian multi-target tracking based on YOLOv3
Li et al. MPAT: Multi-path attention temporal method for video anomaly detection
Gao et al. Lightweight high-precision pedestrian tracking algorithm in complex occlusion scenarios.
Pan et al. Learning to Track by Bi-Directional Long Short-Term Memory Networks
CN117853759B (en) Multi-target tracking method, system, equipment and storage medium
CN113379794B (en) Single-target tracking system and method based on attention-key point prediction model

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20230321

RJ01 Rejection of invention patent application after publication