CN115457082A - Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement - Google Patents

Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement Download PDF

Info

Publication number
CN115457082A
CN115457082A CN202211067913.6A CN202211067913A CN115457082A CN 115457082 A CN115457082 A CN 115457082A CN 202211067913 A CN202211067913 A CN 202211067913A CN 115457082 A CN115457082 A CN 115457082A
Authority
CN
China
Prior art keywords
pedestrian
feature
target
feature fusion
target tracking
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
CN202211067913.6A
Other languages
Chinese (zh)
Inventor
周彦
陈俊宇
王冬丽
杜镇楠
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.)
Xiangtan University
Original Assignee
Xiangtan University
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 Xiangtan University filed Critical Xiangtan University
Priority to CN202211067913.6A priority Critical patent/CN115457082A/en
Publication of CN115457082A publication Critical patent/CN115457082A/en
Pending legal-status Critical Current

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
    • G06T7/248—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving reference images or patches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • 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
    • 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
    • G06T2207/20084—Artificial neural networks [ANN]
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30241—Trajectory

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention discloses a pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement, which integrates three modules of feature extraction, target detection and data association into a network architecture. According to the algorithm, two continuous frames are used as a link node as input, in order to strengthen a resnet50 backbone network, the traditional convolution is replaced by the Incep convolution, and the receptive field of a feature extraction network is strengthened; in order to enhance the adaptability to target deformation, the extracted features are subjected to weighted bidirectional pyramid fusion processing, and more attention is paid to important feature information; in order to enable the network to better process pedestrian targets in crowded scenes, information is enhanced by adding a context-sensitive prediction module into the network and placing a wider and deeper convolution prediction module on the top of layers with different step sizes; and finally, regressing the same target boundary box pair in the adjacent frames, carrying out similarity comparison in the common frame, carrying out data association by using IOU matching, and outputting a tracking track. The algorithm improves the precision of multi-target tracking, and can meet the requirement of the video monitoring field on pedestrian target tracking.

Description

Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement
Technical Field
The invention relates to the technical field of intelligent monitoring, in particular to a pedestrian multi-target tracking algorithm enhanced by multi-feature fusion.
Background
With the continuous improvement of living standard and population growth of people, city monitoring in China faces a plurality of problems, for example, illegal molecules and subsequent tracking may exist in market population, and ensuring the stability of society becomes an important problem for city monitoring departments. Pedestrian tracking is a key basic technology for solving urban safety, has a very rich research value in a real scene, and can provide important information for urban safety by detecting a pedestrian target and tracking pedestrians in follow-up, so that the multi-target tracking of the pedestrians becomes a key problem to be solved for the urban safety.
In the past, multi-target tracking develops rapidly and makes great progress, but most of the existing multi-target tracking field solutions are detection and tracking paradigms based on two-stage models, and the paradigms divide multi-target tracking into two independent tasks, namely detection and association. The method comprises the steps of firstly obtaining a boundary box of an object in each frame through an existing detector, then forming a track through data association cross-frame linking, and embedding an identity to distinguish the objects in an association process. Such a two-step process reveals two ways to improve tracking performance, one by enhancing detection and the other by enhancing data correlation. The tracking process has the influence of shielding, so that objects are overlapped to cause missed detection, and the difficulty of data association is increased. Object detection in successive video frames is difficult to produce stable and reliable results, and the temporal connection between the upper and lower frames is easily lost. The two tasks are separated, although the performance is good, the real-time performance is poor, and the use of an actual scene is not satisfied. Under the conditions of simple model and low calculation cost, the multi-target tracking method based on deep learning can effectively meet the real-time requirement to improve the tracking performance.
Disclosure of Invention
In view of this, the present invention provides a pedestrian multi-target tracking algorithm with multi-feature fusion enhancement to avoid the disadvantages in the prior art, which improves the efficiency of pedestrian multi-target tracking and can meet the requirements in the city monitoring field.
The purpose of the invention is realized by the following technical scheme: the pedestrian multi-target tracking algorithm with the multi-feature fusion enhancement is provided, and comprises the following steps:
s1, performing image preprocessing on a video frame sequence of a pedestrian video set, and extracting characteristic information;
s2, further feature fusion is needed in the extracted feature information, and richer target features are obtained;
and S3, a context sensitive prediction module is used, so that the pedestrian target in a crowded scene can be well processed, and the information is enhanced.
And S4, calculating a regression frame pair of the same target in adjacent frames, performing IOU similarity matching, and outputting a tracking track.
As a further improvement, the image preprocessing in step S1 includes an improvement of extracting feature information, and an inclusion convolution is added to replace a conventional convolution, so as to enhance the sensing region with a more flexible search space, which can improve the search capability of GNN, and enable the model to have an effective receptive field.
As a further improvement, the step improves feature information extraction, and the inclusion convolution is adopted as follows:
Figure BDA0003828014670000021
in the formula (I), the compound is shown in the specification,
Figure BDA0003828014670000022
and
Figure BDA0003828014670000023
is that the filter goes from 1 to d on the x-axis and y-axis of the ith output channel max Expansion of (C) out Is the number of output channels.
As a further improvement, the feature fusion in step S2 includes a weighted bidirectional feature pyramid structure. Firstly, deleting a node with only one input, wherein if one node only has one input boundary, the node has no great effect in the process of feature fusion, and the network can be simplified by deleting redundant nodes; secondly, an edge is directly added to an original input node and an original output node, and more useful feature information is fused under the condition that too much calculation cost is not increased; thirdly, the importance of different features can be effectively learned by adopting multi-scale fusion from top to bottom and from bottom to top through repeated stacking, and higher-level feature fusion is realized.
As a further improvement, the step improves the feature fusion and adopts a weighted bidirectional feature pyramid. Multiscale feature layer
Figure BDA0003828014670000031
Wherein
Figure BDA0003828014670000032
Is represented by i By an operation f of aggregating different features and outputting the new feature
Figure BDA0003828014670000033
Taking layer 6 as an example, the calculation formula is:
Figure BDA0003828014670000034
Figure BDA0003828014670000035
in the formula (I), the compound is shown in the specification,
Figure BDA0003828014670000036
in order to have the middle characteristic from top to bottom,
Figure BDA0003828014670000037
for the bottom-up output features, ε =0.0001 is set to avoid numerical instability during the calculation, and the rest features are calculated in a similar operation manner.
As a further improvement, the context sensitive prediction module is included in the step S3. In crowded scenes, pedestrian targets are not easily distinguished in appearance, and the receptive field is increased by placing wider and deeper convolution prediction modules on top of layers with different step sizes, so that the prediction modules obtain better classification and localization functions and enhance information.
As a further improvement, in step S4, a final tracking track is formed, specifically, the same target regression frame pair in adjacent frames is subjected to IOU matching, and a target with high similarity can be determined as the same target, and the tracking track is output.
Aiming at the research, the invention provides a pedestrian multi-target tracking algorithm enhanced by multi-feature fusion. Firstly, extracting feature information by using Resnet50 based on increment convolution as a backbone network, sending the extracted feature information to a weighted bidirectional feature pyramid to generate a multi-scale feature representation, and then performing context sensitive prediction enhancement operation on the multi-scale feature. In order to associate the same target in adjacent frames, adjacent multi-scale feature maps generated in a backbone network are connected in series, and then are sent into a prediction network to return a bounding box pair. And performing K-means clustering on all ground truth value bounding boxes, and distributing each cluster to a corresponding pyramid layer for subsequent prediction of a specific scale. The prediction network comprises three branches, namely a classification branch, an identity verification branch and a regression block pair regression branch. The classification branch judges whether the region is a target or a background by predicting the confidence coefficient score of the foreground region through the target detection frame, the identity verification branch is used for judging whether the detection frame pair belongs to the same target or not, and the regression branch predicts the coordinates of the same target frame pair. The three branches use the joint attention, the prediction confidence maps of the classification branch and the identity verification branch as the attention maps, so that the attention maps of the two branches can be complemented to achieve more attention-focused and effective information, then the attention maps are multiplied by the combined features and input into the regression branches, the more attention-focused pedestrian targets in the regression process are promoted to avoid the interference of the regression boxes on the irrelevant information in the matching process, and the accuracy of the same target detection box pair is ensured to a certain extent. The algorithm improves the accuracy of pedestrian multi-target tracking, and can be applied to tracking pedestrians in daily scenes.
Drawings
The invention is further illustrated by means of the attached drawings, but the embodiments in the drawings do not constitute any limitation to the invention, and for a person skilled in the art, other drawings can be obtained on the basis of the following drawings without inventive effort.
FIG. 1 is an overall frame diagram of a multi-feature enhanced pedestrian multi-target tracking algorithm;
fig. 2 is a structural diagram of the inclusion convolution.
FIG. 3 is a block diagram of a weighted bidirectional feature pyramid.
Fig. 4 is a block diagram of a context sensitive prediction module.
FIG. 5 is a graph of the tracking effect of the algorithm of the present invention under a fixed shot.
FIG. 6 is a diagram of the tracking effect of the algorithm of the present invention under a moving lens.
Detailed Description
In order to make those skilled in the art better understand the technical solution of the present invention, the following detailed description of the present invention is provided with reference to the accompanying drawings and specific embodiments, and it is to be noted that the embodiments and features of the embodiments of the present application can be combined with each other without conflict.
Referring to fig. 1, which is a general frame diagram of an algorithm, the implementation of the present invention provides a multi-feature enhanced pedestrian multi-target tracking algorithm, which includes the following steps:
s1, performing image preprocessing on a video frame sequence of a pedestrian video set, and extracting characteristic information;
the image preprocessing of the step comprises the improvement of extracting feature information, the inclusion convolution is added to replace the traditional convolution, and the structure diagram of the inclusion convolution is shown in figure 2. The convolution strengthens the perception area by a more flexible search space, and can improve the search capability of the GNN, so that the model can have an effective receptive field.
Specifically, the inclusion convolution space contains multiple modes of dilation, with the dilation of each axis, each channel, and each convolution layer being independent to provide a dense range of effective receptive fields. The inclusion convolution has independent expansions for the two axes in each channel, expressed in the form:
Figure BDA0003828014670000051
in the formula (I), the compound is shown in the specification,
Figure BDA0003828014670000052
and
Figure BDA0003828014670000053
is that the filter goes from 1 to d on the x-axis and y-axis of the ith output channel max Expansion of (C) out Is the number of output channels.
S2, further feature fusion is needed in the extracted feature information, and richer target features are obtained;
the feature fusion of this step includes a weighted bidirectional feature pyramid structure, and the structure diagram of the weighted bidirectional feature pyramid is shown in fig. 3. Firstly, deleting a node with only one input, if one node only has one input boundary, the node has no great effect in the process of feature fusion, and deleting redundant nodes can simplify the network; secondly, an edge is directly added to an original input node and an original output node, and more useful feature information is fused under the condition that too much calculation cost is not increased; thirdly, the importance of different features can be effectively learned by adopting multi-scale fusion from top to bottom and from bottom to top through repeated stacking, and higher-level feature fusion is realized.
Specifically, the step improves feature fusion, and three times of stacking are repeated by adopting a weighted bidirectional feature pyramid. Multiscale feature layer
Figure BDA0003828014670000061
Wherein
Figure BDA0003828014670000062
Represents l i Is characterized by aAggregating the operations f of different features and outputting the new feature
Figure BDA0003828014670000063
Taking layer 6 as an example, the calculation formula is:
Figure BDA0003828014670000064
Figure BDA0003828014670000065
in the formula (I), the compound is shown in the specification,
Figure BDA0003828014670000066
in order to have the middle characteristic from top to bottom,
Figure BDA0003828014670000067
epsilon =0.0001 is set for the bottom-up output feature to avoid numerical instability during calculation, and the rest features are calculated in a similar operation manner.
And S3, a context sensitive prediction module is used, so that the pedestrian target in a crowded scene can be well processed, and the information is enhanced.
This step includes a context sensitive prediction module, and the model architecture can be seen in fig. 4. In crowded scenes, pedestrian targets are not easily distinguished in appearance, and the receptive field is increased by placing wider and deeper convolution prediction modules on top of layers with different step sizes, so that the prediction modules obtain better classification and localization functions and enhance information.
Specifically, a plurality of convolution operations of 3 × 3 and 1 × 1 are added in the module, and the receiving field corresponding to the asynchronous length of the network is increased. By letting the network outputs share the benefits of a wider and deeper network, the prediction module becomes deeper and wider, and better features can be obtained for subsequent subtasks.
And S4, calculating a regression frame pair of the same target in adjacent frames, performing IOU similarity matching, and outputting a tracking track.
The method comprises the steps of forming a final tracking track, specifically performing IOU matching on the same target regression frame pair in adjacent frames, confirming the targets with high similarity as the same target, and outputting the tracking track.
Specifically, only slight changes exist in the boundary boxes of the same target on adjacent frames, the affinity matrix is obtained by calculating the IOU, the optimal matching of the detection boxes of the same target is realized by applying the Hungarian algorithm, and the track where the target is successfully matched is updated.
In summary, the multi-feature enhanced pedestrian multi-target tracking algorithm has the following advantages:
1) In the image preprocessing step, the conventional convolution is replaced by the inclusion convolution, so that the searching capability of the backbone network is improved, and the network can fit the effective receptive field to the video data set.
2) In the image feature fusion step, an efficient weighted bidirectional feature pyramid network is adopted, so that the problems that feature information is easy to lose in the transmission process, targets are identified and positioned disadvantageously and the like are effectively solved, and higher-level feature fusion is realized.
3) In the step of processing the pedestrian target in the crowded scene, a context sensitive prediction module is adopted. Considering that the nature and size of the detected object are easily influenced by environmental factors and the like, particularly under a crowded scene, the differences of pedestrian targets in appearance are not easy to distinguish, the module can well enhance the information, and the problem is solved.
In the description above, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced in other ways than those specifically described herein, and therefore should not be construed as limiting the scope of the present invention.
In conclusion, although the present invention has been described with reference to the preferred embodiments, it should be noted that, although various changes and modifications may be made by those skilled in the art, they should be included in the scope of the present invention unless they depart from the scope of the present invention.

Claims (7)

1. A pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement is characterized by comprising the following steps:
s1, performing image preprocessing on a video frame sequence of a pedestrian video set, and extracting characteristic information;
s2, further feature fusion is needed in the extracted feature information, and richer target features are obtained;
and S3, a context sensitive prediction module is used, so that the pedestrian target in a crowded scene can be well processed, and the information is enhanced.
And S4, calculating a regression frame pair of the same target in adjacent frames, performing IOU similarity matching, and outputting a tracking track.
2. The multi-feature fusion enhanced pedestrian multi-target tracking algorithm as claimed in claim 1, wherein the image preprocessing in the step S1 includes improvement of feature information extraction, and inclusion convolution is added to replace conventional convolution, so that a more flexible search space is used to strengthen a sensing region, thereby improving the search capability of GNN and enabling a model to have an effective receptive field.
3. The pedestrian multi-target tracking algorithm according to claim 2, wherein the step of improving feature information extraction adopts an increment convolution as follows:
Figure FDA0003828014660000011
in the formula (I), the compound is shown in the specification,
Figure FDA0003828014660000012
and
Figure FDA0003828014660000013
is that the filter goes from 1 to d on the x-axis and y-axis of the ith output channel max Expansion of (C) out Is the number of output channels.
4. The multi-feature fusion enhanced pedestrian multi-target tracking algorithm according to claim 1, wherein the feature fusion in the step S2 includes a weighted bidirectional feature pyramid structure. Firstly, deleting a node with only one input, if one node only has one input boundary, the node has no great effect in the process of feature fusion, and deleting redundant nodes can simplify the network; secondly, an edge is directly added to an original input node and an original output node, and more useful feature information is fused under the condition that too much calculation cost is not increased; thirdly, by adopting multi-scale fusion from top to bottom and from bottom to top, the importance of different features can be effectively learned by repeated stacking for many times, and higher-level feature fusion is realized.
5. The pedestrian multi-target tracking algorithm of claim 4, wherein the step of improving feature fusion employs a weighted bidirectional feature pyramid. Multiscale feature layer
Figure FDA0003828014660000021
Wherein
Figure FDA0003828014660000022
Is represented by i By an operation f of aggregating different features and outputting the new feature
Figure FDA0003828014660000023
Taking layer 6 as an example, the calculation formula is:
Figure FDA0003828014660000024
Figure FDA0003828014660000025
in the formula (I), the compound is shown in the specification,
Figure FDA0003828014660000026
in order to have the middle characteristic from top to bottom,
Figure FDA0003828014660000027
for the bottom-up output features, ε =0.0001 is set to avoid numerical instability during the calculation, and the rest features are calculated in a similar operation manner.
6. The multi-feature fusion enhanced pedestrian multi-target tracking algorithm according to claim 1, wherein the step S3 comprises a context sensitive prediction module. In crowded scenes, pedestrian targets are not easily distinguished in appearance, and the receptive field is increased by placing wider and deeper convolution prediction modules on top of layers with different step sizes, so that the prediction modules obtain better classification and localization functions and enhance information.
7. The multi-feature fusion enhanced pedestrian multi-target tracking algorithm according to claim 1, wherein in the step S4, a final tracking track is formed, the same target regression box pair specifically represented in adjacent frames is subjected to IOU matching, and a target with high similarity can be determined as the same target, and the tracking track is output.
CN202211067913.6A 2022-09-01 2022-09-01 Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement Pending CN115457082A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211067913.6A CN115457082A (en) 2022-09-01 2022-09-01 Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211067913.6A CN115457082A (en) 2022-09-01 2022-09-01 Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement

Publications (1)

Publication Number Publication Date
CN115457082A true CN115457082A (en) 2022-12-09

Family

ID=84301808

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211067913.6A Pending CN115457082A (en) 2022-09-01 2022-09-01 Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement

Country Status (1)

Country Link
CN (1) CN115457082A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115620242A (en) * 2022-12-19 2023-01-17 城云科技（中国）有限公司 Multi-line person target re-identification method, device and application
CN117853759A (en) * 2024-03-08 2024-04-09 山东海润数聚科技有限公司 Multi-target tracking method, system, equipment and storage medium
CN117853759B (en) * 2024-03-08 2024-05-10 山东海润数聚科技有限公司 Multi-target tracking method, system, equipment and storage medium

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115620242A (en) * 2022-12-19 2023-01-17 城云科技（中国）有限公司 Multi-line person target re-identification method, device and application
CN117853759A (en) * 2024-03-08 2024-04-09 山东海润数聚科技有限公司 Multi-target tracking method, system, equipment and storage medium
CN117853759B (en) * 2024-03-08 2024-05-10 山东海润数聚科技有限公司 Multi-target tracking method, system, equipment and storage medium

Similar Documents

Publication Publication Date Title
Dai et al. TIRNet: Object detection in thermal infrared images for autonomous driving
CN111259786B (en) Pedestrian re-identification method based on synchronous enhancement of appearance and motion information of video
Yuan et al. Anomaly detection in traffic scenes via spatial-aware motion reconstruction
CN112560656B (en) Pedestrian multi-target tracking method combining attention mechanism end-to-end training
CN107145862B (en) Multi-feature matching multi-target tracking method based on Hough forest
CN110288627B (en) Online multi-target tracking method based on deep learning and data association
CN105069434B (en) A kind of human action Activity recognition method in video
CN111862145B (en) Target tracking method based on multi-scale pedestrian detection
Zhang et al. Coarse-to-fine object detection in unmanned aerial vehicle imagery using lightweight convolutional neural network and deep motion saliency
CN115830075A (en) Hierarchical association matching method for pedestrian multi-target tracking
CN110619268A (en) Pedestrian re-identification method and device based on space-time analysis and depth features
Juyal et al. Deep learning methods for object detection in autonomous vehicles
CN105957103B (en) A kind of Motion feature extraction method of view-based access control model
CN113920472A (en) Unsupervised target re-identification method and system based on attention mechanism
Jemilda et al. Moving object detection and tracking using genetic algorithm enabled extreme learning machine
CN113808166B (en) Single-target tracking method based on clustering difference and depth twin convolutional neural network
CN115457082A (en) Pedestrian multi-target tracking algorithm based on multi-feature fusion enhancement
Kim et al. Novel on-road vehicle detection system using multi-stage convolutional neural network
Ji et al. A hybrid model of convolutional neural networks and deep regression forests for crowd counting
Li et al. Real-time tracking algorithm for aerial vehicles using improved convolutional neural network and transfer learning
CN116721398A (en) Yolov5 target detection method based on cross-stage route attention module and residual information fusion module
CN111291785A (en) Target detection method, device, equipment and storage medium
Zhang et al. Bus passenger flow statistics algorithm based on deep learning
CN116245913A (en) Multi-target tracking method based on hierarchical context guidance
Badal et al. Online multi-object tracking: multiple instance based target appearance 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