CN111814604A - Pedestrian tracking method based on twin neural network - Google Patents
Pedestrian tracking method based on twin neural network Download PDFInfo
- Publication number
- CN111814604A CN111814604A CN202010584083.9A CN202010584083A CN111814604A CN 111814604 A CN111814604 A CN 111814604A CN 202010584083 A CN202010584083 A CN 202010584083A CN 111814604 A CN111814604 A CN 111814604A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- pedestrians
- short
- neural network
- 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
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 59
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 39
- 230000033001 locomotion Effects 0.000 claims description 53
- 239000011159 matrix material Substances 0.000 claims description 28
- 238000001514 detection method Methods 0.000 claims description 21
- 238000012360 testing method Methods 0.000 claims description 14
- 230000008569 process Effects 0.000 claims description 7
- 238000005516 engineering process Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000012544 monitoring process Methods 0.000 claims description 4
- 239000004576 sand Substances 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 2
- 230000001105 regulatory effect Effects 0.000 claims description 2
- 238000002474 experimental method Methods 0.000 description 10
- 230000006870 function Effects 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013100 final test Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000011282 treatment Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- 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
- G06T2207/30232—Surveillance
-
- 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
Abstract
The invention belongs to the field of computer vision, and particularly relates to a pedestrian tracking method based on a twin neural network, which comprises the following steps: inputting a video; marking the pedestrian; acquiring a pedestrian space-time group; establishing and training a twin neural network, and storing the trained twin neural network; acquiring a short track of a pedestrian; and acquiring a long track of the pedestrian. The method is adopted to track the pedestrians, and the accuracy rate of tracking the pedestrians is effectively improved.
Description
Technical Field
The invention belongs to the field of computer vision, and particularly relates to a pedestrian tracking method based on a twin neural network.
Background
With the development of artificial intelligence technology, computer vision has been widely applied to human daily life such as smart home, video surveillance and intelligent transportation, and pedestrian tracking is one of the key problems in these fields. Because the pedestrian target can change in posture, size, appearance and pedestrian shelter from in the motion process for it becomes very difficult to realize the accurate tracking of pedestrian.
The early target tracking algorithm is mainly based on target modeling or tracking target features, and the main methods are as follows: (1) the feature matching method comprises the steps of firstly extracting target features, and then finding the most similar features in subsequent frames to carry out target positioning; (2) based on a search method, people add a prediction algorithm into tracking and search targets near predicted values, so that the search range is reduced, however, the methods no longer meet the current pedestrian tracking requirements, and new methods are urgently needed to replace the methods; and the traditional method is greatly influenced by factors such as image illumination change, pedestrian posture change, image noise and the like. With the rapid development of deep learning, the method plays an important promoting role in the field of pedestrian tracking, wherein the twin neural network is a model capable of judging whether two input modes are similar or not, and can be used in the field of pedestrian tracking to judge whether appearances of pedestrians are similar or not. Accordingly, there is a need for improvements in the art.
Disclosure of Invention
The invention aims to solve the problem of providing a pedestrian motion trajectory calculation method based on a twin neural network, which is used for tracking pedestrians and effectively improves the accuracy of pedestrian tracking.
In order to solve the above problems, the present invention provides a pedestrian tracking method based on a twin neural network, comprising the following steps:
step 1, video input:
acquiring by monitoring equipment (a camera and other monitoring equipment) outdoors or indoors to obtain a video file, and inputting the acquired video file into a computer, wherein the acquired video file comprises each frame of video image of a video sequence;
step 2, pedestrian marking:
for each frame of video image input in the step 1, detecting and marking the position of a pedestrian on each frame of video image by using a DPM pedestrian detection technology to obtain a video sequence with a pedestrian position mark;
step 3, acquiring pedestrian space-time groups:
marking the position with the pedestrian in the step 2The recorded video sequence is divided into n sections at intervals of 1 second, and each section is a video image i; for each video image i, obtaining pedestrian space-time group by using a hierarchical clustering method according to pedestrian positionsN is a positive integer, i is not more than n, and j represents the space-time group serial number in the ith video image;
step 4, establishing and training a twin neural network, and storing the trained twin neural network;
step 5, acquiring a short track of the pedestrian:
step 5-1, for each spatio-temporal group in each video image i obtained in step 3The pedestrians in the pedestrian seat are combined in pairs, and each combination is two pictures containing the pedestrians and having the size of 128 × 64 × 3;
step 5-2, stacking the two pictures containing the pedestrians in each combination obtained in the step 5-1 and having the size of 128 × 64 × 3 into a form of 128 × 64 × 6, inputting the pictures into the twin neural network trained in the step 4, and obtaining each space-time groupThe appearance similarity P of each pair of pedestrians in the interiorAS;
Step 5-3, calculating the pedestrian movement speed
For the pedestrian position p in the t-th frame, finding the detection p closest to p by calculating the distance in the adjacent framekK is the index of the frame, and each space-time group is calculated according to the following formulaInner pedestrian velocity vpedestrian:
Step 5-4, calculating the movement affinity
Calculating each spatiotemporal group according to the following formula by using the pedestrian movement speed calculation obtained in the step 5-2Within each pedestrian pair motion affinity PMA:
PMA=max(1-β*e(P1,P2),0),
e(P1,P2)=min(eforward(P1,P2),ebackward(P1,P2)),
Wherein P is1And P2For two pedestrians, β is the parameter to be regulated, eXforward,eYforward,eXbackwardAnd eYbackwardIs the forward error and the reverse error of the pedestrian movement in the X direction and the Y direction, eforwardAnd ebackwardIs the forward error and reverse error of the pedestrian's motion, P1x,P1y,P2xAnd P2yIs the X and Y coordinates of two pedestrian positions, t1And t2Is the frame index of two pedestrians, v obtained in step 5-3pedestrianDecomposed into X-and Y-direction movement velocities, v1x,v1y,v2xAnd v2yThe moving speeds of the two pedestrians in the X direction and the Y direction respectively;
step 5-5, establishing a pedestrian correlation matrix
Using sigmoid function to combine each spatiotemporal group obtained in step 5-2Appearance similarity P of pedestrians in each teamASAnd each spatio-temporal group obtained in step 5-4Within each pedestrian pair motion affinity PMAThe pedestrian correlation matrix PeC is fused as follows:
step 5-6, solving classification
Using the maximum correlation principle for each spatio-temporal group obtained in step 5-5Solving and classifying the pedestrian correlation matrix PeC in the pedestrian identification matrix, wherein each obtained class respectively represents the pedestrian with the same identity; a graph G ═ V, E, W is provided, where the vertex V represents each pedestrian detection, E represents the side to which the pedestrian detections are connected, and W represents the correlation between two pedestrians, and the maximum correlation principle is as follows:
wherein v is1,v2,v3Representing pedestrian detection in V, w represents the correlation degree between pedestrians, and the value of x is 0 or 1;
step 5-7, obtaining short tracks of pedestrians
Merging the pedestrians in each class obtained in the step 5-6, so as to obtain short tracks of the pedestrians in each video image i;
step 6, acquiring long tracks of pedestrians
6-1, selecting key pedestrians
Processing short tracks of the pedestrians in each section of the video image i obtained in the step 5-5 in the sliding time window, and selecting key pedestrians in the short tracks in each section of the video image i by using the appearance similarity of the pedestrians for all the short tracks in one sliding time window, and using the key pedestrians to represent one short track;
step 6-2, combining every two key pedestrians in short tracks
In a sliding time window, combining every two key pedestrians with each short track in a certain section of video image i selected in the step 6-1 and every two key pedestrians with each short track in other sections of video images j, wherein each combination is two pictures with the size of 128 × 64 × 3 and containing the key pedestrians;
step 6-3, calculating the similarity of the short tracks
Inputting two pictures with the size of 128 × 64 × 3 containing key pedestrians into the twin neural network trained in the step 4, and outputting the result as the appearance similarity of the two key pedestrians with short tracksThen, the operation is performed according to the following formula, and the similarity of the two short tracks is output:
wherein the content of the first and second substances,representing the appearance similarity of two short-track key pedestrians, and N representing the coincidence conditionK ofASSet, N represents the number of elements in N;
step 6-4, calculating the movement speed v of each short tracktracklet
The motion speed of each short track is calculated according to the following formula:
PSindicating the start of a short track, PeIndicating the end of a short track, TSIndicating the start of a short track, TeA termination frame representing a short track;
step 6-5, calculating the motion affinity Tr between the short tracksMA
TrMA=max(1-β*e(Tr1,Tr2),0)
e(Tr1,Tr2)=eforward(Tr1,Tr2)+ebackward(Tr1,Tr2)
Wherein P is1And P2For two short tracks, beta is the parameter to be adjusted, eXforward,eYforward,eXbackwardAnd eYbackwardIs the forward and reverse error of the short track in the X and Y directions, eforwardAnd ebackwardAre the forward error and the reverse error of the short track motion,andis the end point coordinate of a short track,andis the starting point coordinate of another short track, t1And t2Of one short track end point coordinate and another short trackFrame index of start point coordinates, v obtained by S0604trackletDecomposed into X-and Y-direction movement velocities, v1x,v1y,v2xAnd v2yThe moving speeds of the two short tracks in the X direction and the Y direction are respectively;
step 6-6, establishing a short track correlation matrix
Similarity Tr of short tracks obtained in step 6-3 by using sigmoid functionsAnd the kinematic affinity Tr of the short trajectory obtained in step 6-5MAAnd (3) fusing into a short track correlation matrix TrC:
step 6-7, solving classification
Solving and classifying the short track correlation matrix TrC obtained in the step 6-6 by using a maximum correlation principle, wherein each obtained class respectively represents the short tracks with the same identity, the maximum correlation principle is as follows, and a graph G ═ V, E, W is set, wherein a vertex V represents each short track, E represents an edge connected with the short tracks, and W represents the correlation between the two short tracks:
wherein v is1,v2,v3Representing short tracks in V, w represents the correlation degree between the short tracks, and the value of x is 0 or 1;
step 6-8, merging the growth tracks
Merging the short tracks of each type obtained by solving and classifying in the step 6-7, so as to merge the short tracks of the pedestrians with the same identity into a new track; as the sliding time window progresses, a complete long trajectory for all pedestrians is finally obtained.
The invention relates to an improvement of a pedestrian tracking method based on a twin neural network, which comprises the following steps: step 4, the establishment and training of the twin neural network comprises the following steps:
step 4-1, creating a twin neural network
A twin neural network is built by using a Pythrch frame, and the structure is as follows: two 128 × 64 × 3 pedestrian pictures are stacked into 128 × 64 × 6 and then used as the input of the twin neural network, and pass through three layers of convolution layers, wherein the convolution kernel size of the first layer is 9 × 12, the convolution kernel size of the second layer is 5 × 16, and the convolution kernel size of the third layer is 5 × 24; each layer convolution contains batch normalization, relu activation function, pooling of 2 x 2; the output of the three-layer convolution is subjected to one-dimensional operation to obtain a vector of 1 x 1152, then the vector passes through two full-connected layers with the sizes of 1 x 150 and 1 x 2, and finally the output result is the appearance similarity of two pedestrian pictures;
step 4-2, training twin neural network
Data set setting: downloading a CUHK02 pedestrian tracking data set from the Internet, wherein the data set comprises 1816 pedestrian identities, each pedestrian is shot by 2 camera angles, each camera angle has 2 photos, the first 80% of pedestrians are taken as a training set, and the rest pedestrians are taken as a testing set;
training process: two rounds of twin neural network training established in the step 4-1 are carried out, in the first round, training data are pictures of a first shooting angle of pedestrians in a training set, sample combinations are randomly generated in the training process, and the proportion of positive samples to negative samples is controlled to be 1: 1, the positive samples are two pedestrians with the same identity, and the negative samples are two pedestrians with different identities; in the second round, the picture of the second camera angle of the pedestrian in the training set is used as training data, and the generation mode of the positive sample and the negative sample is the same as that of the first round;
training and testing: after training is finished, testing the trained twin neural network by using the test set, wherein the output of the twin neural network is the appearance similarity of two pedestrians, and when the appearance similarity of the two pedestrians with the same identity is more than or equal to 0.5, the two pedestrians are considered to be correct; on the contrary, the two pedestrians with different identities are considered to be correct when the appearance similarity is less than or equal to 0.5; and finally, the accuracy of the test result reaches more than 98%, and the trained twin neural network is stored.
The invention has the following technical advantages:
1. according to the invention, the twin neural network is used for obtaining the appearance similarity between pedestrians, the influence of factors such as image illumination change, pedestrian posture change and image noise is small, and accurate pedestrian tracking can be still carried out when the problems are met;
2. the invention realizes pedestrian tracking in a layered mode: firstly, short tracks are generated by pedestrian detection, then long tracks are generated by the short tracks, and the generation of the short tracks and the long tracks is realized by the twin neural network in the two steps;
3. the pedestrian information is fused based on the sigmoid function, and the pedestrian (short track) similarity and the motion affinity mapping based on the twin neural network are fused into the pedestrian correlation (short track correlation);
4. according to the invention, the appearance similarity and the motion affinity of the pedestrian are fully considered when the short track of the pedestrian is generated, the similarity and the motion affinity of the short track are fully considered when the long track of the pedestrian is generated, each step is comprehensively considered from two aspects, so that the tracking failure caused by the shielding of the pedestrian can be effectively avoided, and in addition, each step is solved and classified according to the maximum correlation principle to ensure the correct tracking of the pedestrian.
The invention is derived from the following fund: zhejiang province natural science fund project (LQ19F030014), and Zhejiang university youth innovation project (2019Q 035).
Drawings
The following detailed description of embodiments of the invention is provided in conjunction with the appended drawings:
FIG. 1 is a schematic diagram of an algorithm flow of a twin neural network-based pedestrian tracking method;
fig. 2 is a pedestrian position image in the PETS2009-S2L1 dataset detected using the DPM detector in step 2 of experiment 1;
fig. 3 is an image of the pedestrian location in the Town Center data set detected using the DPM detector in step 2 of experiment 1;
fig. 4 is a schematic diagram of an error between the trajectory of a pedestrian in the PETS2009-S2L1 data set obtained in step 5 of experiment 1 and the trajectory noted in the groudtuth file;
FIG. 5 is a schematic illustration of the error between the pedestrian's trajectory in the Town Center dataset obtained in step 5 of experiment 1 and the trajectory noted in the grountruth file;
FIG. 6 is a graph of the experimental effect of pedestrian trajectories for the PETS2009-S2L1 dataset obtained at step 5 in experiment 1;
fig. 7 is a graph showing the experimental effect of the pedestrian trajectory of the Town Center data set obtained in step 5 of experiment 1.
Detailed Description
The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto.
Embodiment 1, a pedestrian tracking method based on a twin neural network, as shown in fig. 1, includes the following steps:
s01, video input
Inputting a video file containing an object to be tracked into a computer, wherein the video file comprises each frame of video image, and the video file is acquired by a camera and other monitoring equipment outdoors or indoors;
s02, pedestrian marking:
for each frame of video image input in the step S01, detecting and marking the position of a pedestrian on each frame of video image by using a DPM pedestrian detection technology, to obtain a video sequence with a pedestrian position mark;
the DPM pedestrian detection technology belongs to the conventional technology, and for example, the DPM pedestrian detection algorithm disclosed in IEEETransactions on Pattern Analysis & Machine Analysis, 2010, by Felzenszwalb et al can be referred to.
S03, acquiring pedestrian space-time groups:
splitting the video sequence with the pedestrian position mark in S02 into n sections every 1 second, wherein each section is a video image i;
for each video image i, a plurality of pedestrian space-time groups are obtained by using a hierarchical clustering method according to the positions of pedestriansN is a positive integer, i is not more than n, and j represents the space-time group serial number in the ith video image;
s04, creating and training twin neural network
S0401, creation of twin neural networks:
a twin neural network is built by using a Pythrch frame, and the structure is as follows: two 128 × 64 × 3 pedestrian pictures are stacked into 128 × 64 × 6 and then used as the input of the twin neural network, and pass through three layers of convolution layers, wherein the convolution kernel size of the first layer is 9 × 12, the convolution kernel size of the second layer is 5 × 16, and the convolution kernel size of the third layer is 5 × 24. Each layer convolution contains batch normalization, relu activation function, pooling of 2 x 2; the output of the three-layer convolution is subjected to one-dimensional operation to obtain a vector of 1 x 1152, then the vector passes through two full-connected layers with the sizes of 1 x 150 and 1 x 2, and finally the output result is the appearance similarity of two pedestrian pictures;
s0402, training twin neural network
Data set setting: downloading a CUHK02 pedestrian tracking data set from the Internet, wherein the data set comprises 1816 pedestrian identities, each pedestrian is shot by 2 camera angles, each camera angle has 2 photos, the first 80% of pedestrians are taken as a training set, and the rest pedestrians are taken as a testing set;
training process: two rounds of twin neural network training established for S0401 are performed, in the first round, training data are pictures of a first shooting angle of pedestrians in a training set, sample combinations are randomly generated in the training process, and the proportion of positive samples to negative samples is controlled to be 1: 1, the positive samples are two pedestrians with the same identity, and the negative samples are two pedestrians with different identities; in the second round, the picture of the second camera angle of the pedestrian in the training set is used as training data, and the generation mode of the positive sample and the negative sample is the same as that of the first round;
training and testing: after training is finished, testing the trained twin neural network by using the test set, wherein the output of the twin neural network is the appearance similarity of two pedestrians, and when the appearance similarity of the two pedestrians with the same identity is more than or equal to 0.5, the two pedestrians are considered to be correct; on the contrary, the two pedestrians with different identities are considered to be correct when the appearance similarity is less than or equal to 0.5; the accuracy of the final test result reaches more than 98%, and the trained twin neural network is stored;
s05, acquiring short tracks of pedestrians
For each video image i obtained in S03 and the pedestrian space-time group thereinThe following treatments were carried out:
s0501, grouping each space-timeThe pedestrians in the pedestrian seat are combined in pairs, and each combination is two pictures containing the pedestrians and having the size of 128 × 64 × 3;
s0502, stacking two pictures with the size of 128 × 64 × 3 containing pedestrians in each combination obtained by S0501 into a form of 128 × 64 × 6, inputting the pictures into the S0402 trained twin neural network, and obtaining each space-time groupThe appearance similarity P of each pair of pedestrians in the interiorAS;
S0503, calculating the pedestrian movement speed:
the pedestrian position in the t frame is p, and a plurality of detections p with the closest distance to p are found in the adjacent frames by calculating the distancek(k is the index of the frame) for each spatio-temporal groupInner pedestrian velocity vpedestrianCalculated as follows:
s0504, calculation of kinetic affinity
Pedestrian movement velocity v obtained using S0502pedestrianCalculating each space-time group according to the following formulaWithin each pedestrian pair motion affinity PMA:
PMA=max(1-β*e(P1,P2),0),
e(P1,P2)=min(eforward(P1,P2),ebackward(P1,P2)),
Wherein P is1And P2For two pedestrians, β is the parameter that can be adjusted, eXforward,eYforward,eXbackwardAnd eYbackwardIs the forward error and the reverse error of the pedestrian movement in the X direction and the Y direction, eforwardAnd ebackwardIs the forward error and reverse error of the pedestrian's motion, P1x,P1y,P2xAnd P2yIs the X and Y coordinates of two pedestrian positions, t1And t2Is the frame index of two pedestrians, v obtained by S0503pedestrianCan be decomposed into X-direction and Y-direction moving speed, v1x,v1y,v2xAnd v2yThe moving speeds of the two pedestrians in the X direction and the Y direction respectively;
s0505, establishing pedestrian correlation matrix
Using sigmoid function to acquire each spatio-temporal group of S0502Appearance similarity P of pedestrians in each teamASAnd each spatio-temporal group obtained in S0504Within each pedestrian pair motion affinity PMAThe pedestrian correlation matrix PeC is fused as follows:
s0506, solve classification
Each spatio-temporal group obtained for S0505 using the maximum correlation principleSolving and classifying the pedestrian correlation matrix PeC in the pedestrian identification matrix, wherein each obtained class represents the pedestrian with the same identity; a graph G ═ V, E, W is provided, where the vertex V represents each pedestrian detection, E represents the side to which the pedestrian detections are connected, and W represents the correlation between two pedestrians, and the maximum correlation principle is as follows:
wherein v is1,v2,v3Representing pedestrian detection in V, w represents the correlation degree between pedestrians, and the value of x is 0 or 1;
s0507, obtaining short track
Merging the pedestrians in each class obtained in the step S0506, namely obtaining short tracks of the pedestrians in each video image i;
s06, acquiring long tracks of pedestrians
In the sliding time window, the short trajectory of the pedestrian in each segment of video image i obtained in S0507 is processed:
s0601, selecting key pedestrians
For all short tracks in a sliding time window, selecting key pedestrians in the short tracks in each image video i by using the appearance similarity of the pedestrians, and representing one short track by using the key pedestrians;
s0602, two-by-two combination of key pedestrians between short tracks
In a sliding time window, combining every two key pedestrians with each short track in a certain section of video image i selected in S0601 and every two key pedestrians with each short track in other sections of video images j, wherein each combination is two pictures with the size of 128 × 64 × 3 and containing the key pedestrians;
s0603, calculating similarity of short tracks
Inputting two pictures with size of 128 × 64 × 3 containing key pedestrians into the S0402 trained twin neural network in the same way as the S0502, and outputting the result as the appearance similarity of the two key pedestrians with short tracksFurther operation is carried out according to the following formula, and the similarity Tr of the two short tracks is outputs:
Wherein the content of the first and second substances,representing the appearance similarity of two short-track key pedestrians, and N representing the coincidence conditionK ofASSet, N represents the number of elements in N;
s0604, calculating motion speed v of short tracktracklet
The motion speed of each short track is calculated according to the following formula:
PSindicating the start of a short track, PeIndicating the end of a short track, TSIndicating the start of a short track, TeA termination frame representing a short track;
s0605, calculating short-track motion affinity TrMA
Using the moving velocity of the short trajectories obtained in S0604, the moving affinity Tr between the short trajectories was calculated as followsMA
TrMA=max(1-β*e(Tr1,Tr2),0)
e(Tr1,Tr2)=eforward(Tr1,Tr2)+ebackward(Tr1,Tr2)
Wherein P is1And P2For two short tracks, beta is a parameter which can be adjusted, eXforward,eYforward,eXbackwardAnd eYbackwardIs the forward and reverse error of the short track in the X and Y directions, eforwardAnd ebackwardAre the forward error and the reverse error of the short track motion,andis the end point coordinate of a short track,andis anotherCoordinates of the start of the short track of the bar, t1And t2V is obtained from the frame index of the end point coordinate of one short track and the frame index of the start point coordinate of the other short track, S0604trackletCan be decomposed into X-direction and Y-direction moving speed, v1x,v1y,v2xAnd v2yThe moving speeds of the two short tracks in the X direction and the Y direction are respectively;
s0606, establishing short track correlation matrix
Similarity Tr of short tracks obtained by S0603 using sigmoid functionsAnd the kinematic affinity Tr of the short trajectory obtained in S0605MAAnd (3) fusing into a short track correlation matrix TrC:
where λ and μ are adjustable parameters;
s0607, solving classification
Solving and classifying the short track correlation matrix TrC obtained by S0606 by using a maximum correlation principle, wherein each obtained class represents a short track of a pedestrian with the same identity, and a graph G (V, E, W) is set, wherein a vertex V represents each short track, E represents an edge connected with the short tracks, W represents the correlation between the two short tracks, and the maximum correlation principle is as follows:
wherein v is1,v2,v3Representing short tracks in V, w represents the correlation degree between the short tracks, and the value of x is 0 or 1;
s0608, merging growth trajectories
Merging the short tracks of each type obtained by solving and classifying in the S0607, namely merging the short tracks of pedestrians with the same identity into a new track; as the sliding time window progresses, a complete long trajectory for all pedestrians is finally obtained.
Experiment 1, the pedestrian tracking method based on the twin neural network established in the embodiment 1 is used for online tracking:
step 1, video input
Selecting a video data set file of an object to be tracked and inputting the video data set file into a computer, wherein the video data set file comprises each frame of video image of a video sequence and a grountruth file, and the grountruth file comprises a pedestrian track marked artificially;
step 2, pedestrian marking:
for each frame of video image input in step 1, detecting and marking the position of a pedestrian on each frame of video image by using a DPM pedestrian detector, as shown in fig. 2 and 3, obtaining a video sequence with a pedestrian position mark in a PETS2009-S2L1 data set and a TownCenter data set;
step 3, acquiring pedestrian space-time groups:
splitting the video sequence with the pedestrian position marks in the step 2 into n sections every 1 second, wherein each section is a video image i; for each video image i, a plurality of pedestrian space-time groups are obtained by using a hierarchical clustering method according to the positions of pedestriansN is a positive integer, i is not more than n, and j represents the space-time group serial number in the ith video image;
step 4, acquiring short tracks of pedestrians
Step 4-1, for each spatio-temporal group in each video image i obtained in step 3The pedestrians in each group were combined two by two, each group was two pictures with size of 128 x 64 x 3 containing pedestrians, and the two pictures with size of 128 x 64 x 3 in each group were stacked into 128 x 64 x 6 and inputted into the example 1 for trainingTo obtain each spatio-temporal groupThe appearance similarity P of each pair of pedestrians in the interiorAS;
Step 4-2, calculating the pedestrian movement speed:
for the pedestrian position p in the t-th frame, finding several detections p closest to p by calculating the distance in the adjacent framesk(k is the index of the frame) and the pedestrian movement velocity v in each space-time group is calculated according to the following formulapedestrian:
Step 4-3, calculating the movement affinity
Calculating each spatiotemporal group according to the following formula by using the pedestrian movement speed calculation obtained in the step 4-2Within each pedestrian pair motion affinity PMA:
PMA=max(1-β*e(P1,P2),0)
e(P1,P2)=min(eforward(P1,P2),ebackward(P1,P2))
Step 4-4, establishing a pedestrian correlation matrix
Using sigmoid function to combine each spatiotemporal group obtained in step 4-1Appearance similarity P of pedestrians in each teamASAnd each spatio-temporal group obtained in step 4-3Within each pedestrian pair motion affinity PMAThe pedestrian correlation matrix PeC is fused as follows:
step 4-5, solving the classification
Using the maximum correlation principle for each spatio-temporal group obtained in step 4-4Solving and classifying the pedestrian correlation matrix PeC in the pedestrian identification matrix, wherein each obtained class respectively represents the pedestrian with the same identity; a graph G ═ V, E, W is provided, where the vertex V represents each pedestrian detection, E represents the side to which the pedestrian detections are connected, and W represents the correlation between two pedestrians, and the maximum correlation principle is as follows:
step 4-6, obtaining short tracks of pedestrians
Merging the pedestrians in each class, and obtaining short tracks of the pedestrians in each video image i;
step 5, acquiring long tracks of pedestrians
Step 5-1, selecting key pedestrians
For all short tracks in a sliding time window, selecting key pedestrians in the short tracks in each image video i by using the appearance similarity of the pedestrians, and using the pedestrians to represent one short track;
step 5-2, combining every two key pedestrians in short tracks
In a sliding time window, combining every two key pedestrians with each short track in a certain section of video image i selected in the step 5-1 and every two key pedestrians with each short track in other sections of video images j, wherein each combination is two pictures with the size of 128 × 64 × 3 and containing the key pedestrians;
step 5-3, calculating the similarity of the short tracks
Inputting two pictures with the size of 128 × 64 × 3 containing key pedestrians into the twin neural network trained in the embodiment 1, wherein the output result of the trained twin neural network is the appearance similarity of two key pedestrians with short tracksThen, the similarity Tr of the two short tracks is further calculated according to the following formulas:
Step 5-4, calculating the motion speed v of each short tracktracklet
The motion speed of each short track is calculated according to the following formula:
step 5-5, calculating the motion affinity Tr between the short tracksMA
Calculating the motion affinity Tr between the short tracks by using the motion velocity of the short tracks obtained in step 5-4MA
TrMA=max(1-β*e(Tr1,Tr2),0)
e(Tr1,Tr2)=eforward(Tr1,Tr2)+ebackward(Tr1,Tr2)
Step 5-6, establishing a short track correlation matrix
Similarity Tr of short tracks obtained in step 5-3 by using sigmoid functionsAnd the kinematic affinity Tr of the short trajectory obtained in step 5-5MAAnd (3) fusing into a short track correlation matrix TrC:
step 5-7, solving and classifying
Solving and classifying the short track correlation matrix TrC obtained in the step 5-6 by using a maximum correlation principle, wherein each obtained class respectively represents the short tracks with the same identity, the maximum correlation principle is as follows, and a graph G ═ V, E, W is set, wherein a vertex V represents each short track, E represents an edge connected with the short tracks, and W represents the correlation between the two short tracks:
step 5-8, combining the short tracks in each class obtained by solving and classifying in the step 5-7, namely combining the short tracks of pedestrians with the same identity into a new track; as shown in fig. 6 and 7, a complete long trajectory of all pedestrians is finally obtained as the sliding time window goes by.
Comparing the complete long track of the pedestrian obtained in the experiment 1 with the pedestrian track marked artificially in the grountruth file to obtain the final experimental error and experimental effect, as shown in fig. 4 and 5, it is shown that the algorithm can completely accurately and quickly mark the motion track of the pedestrian and highly coincide with the actual track, wherein the pedestrian tracking accuracy of the PETS2009-S2L1 data set reaches 93.23%, and the pedestrian tracking accuracy of the Town Center data set reaches 78.85%; similarly, the invention uses the steps of experiment 1 to verify batch data, and the result is also valid; thus, the procedure of experiment 1 also verifies that the same procedure is also applicable to other on-line traces.
Finally, it is also noted that the above-mentioned lists merely illustrate a few specific embodiments of the invention. It is obvious that the invention is not limited to the above embodiments, but that many variations are possible. All modifications which can be derived or suggested by a person skilled in the art from the disclosure of the present invention are to be considered within the scope of the invention.
Claims (4)
1. A pedestrian tracking method based on a twin neural network is characterized by comprising the following steps:
step 1, video input:
acquiring a video file outdoors or indoors by monitoring equipment, and inputting the acquired video file into a computer, wherein the acquired video file comprises each frame of video image of a video sequence;
step 2, pedestrian marking:
for each frame of video image input in the step 1, detecting and marking the position of a pedestrian on each frame of video image by using a DPM pedestrian detection technology to obtain a video sequence with a pedestrian position mark;
step 3, acquiring pedestrian space-time groups:
splitting the video sequence with the pedestrian position marks in the step 2 into n sections every 1 second, wherein each section is a video image i; for each video image i, obtaining pedestrian space-time group by using a hierarchical clustering method according to pedestrian positionsWherein n is a positive integer, i is not more than n, jRepresenting the spatio-temporal group number in the ith video image;
step 4, establishing and training a twin neural network, and storing the trained twin neural network;
step 5, acquiring a short track of the pedestrian;
and 6, acquiring a long track of the pedestrian.
2. The twin neural network-based pedestrian tracking method according to claim 1, further characterized in that the step 5 comprises the steps of:
step 5-1, for each spatio-temporal group in each video image i obtained in step 3The pedestrians in the pedestrian seat are combined in pairs, and each combination is two pictures containing the pedestrians and having the size of 128 × 64 × 3;
step 5-2, stacking the two pictures containing the pedestrians in each combination obtained in the step 5-1 and having the size of 128 × 64 × 3 into a form of 128 × 64 × 6, inputting the pictures into the twin neural network trained in the step 4, and obtaining each space-time groupThe appearance similarity P of each pair of pedestrians in the interiorAS;
Step 5-3, calculating the pedestrian movement speed
For the pedestrian position p in the t-th frame, finding the detection p closest to p by calculating the distance in the adjacent framekK is the index of the frame, and each space-time group is calculated according to the following formulaInner pedestrian velocity vpedestrian:
Step 5-4, calculating the movement affinity
Calculating each spatiotemporal group according to the following formula by using the pedestrian movement speed calculation obtained in the step 5-2Within each pedestrian pair motion affinity PMA:
PMA=max(1-β*e(P1,P2),0),
e(P1,P2)=min(eforward(P1,P2),ebackward(P1,P2)),
Wherein P is1And P2For two pedestrians, β is the parameter to be regulated, eXforward,eYforward,eXbackwardAnd eYbackwardIs the forward error and the reverse error of the pedestrian movement in the X direction and the Y direction, eforwardAnd ebackwardIs the forward error and reverse error of the pedestrian's motion, P1x,P1y,P2xAnd P2yIs the X and Y coordinates of two pedestrian positions, t1And t2Is the frame index of two pedestrians, v obtained in step 5-3pedestrianDecomposed into X-and Y-direction movement velocities, v1x,v1y,v2xAnd v2yThe moving speeds of the two pedestrians in the X direction and the Y direction respectively;
step 5-5, establishing a pedestrian correlation matrix
Using sigmoid function to combine each spatiotemporal group obtained in step 5-2Appearance similarity P of pedestrians in each teamASAnd each spatio-temporal group obtained in step 5-4Within each pedestrian pair motion affinity PMAThe pedestrian correlation matrix PeC is fused as follows:
step 5-6, solving classification
Using the maximum correlation principle for each spatio-temporal group obtained in step 5-5Solving and classifying the pedestrian correlation matrix PeC in the pedestrian identification matrix, wherein each obtained class respectively represents the pedestrian with the same identity; a graph G ═ V, E, W is provided, where the vertex V represents each pedestrian detection, E represents the side to which the pedestrian detections are connected, and W represents the correlation between two pedestrians, and the maximum correlation principle is as follows:
wherein v is1,v2,v3Representing pedestrian detection in V, w represents the correlation degree between pedestrians, and the value of x is 0 or 1;
step 5-7, obtaining short tracks of pedestrians
And merging the pedestrians in each class obtained in the step 5-6, so as to obtain the short track of the pedestrian in each video image i.
3. The twin neural network-based pedestrian tracking method according to claim 2, further characterized in that the step 6 comprises the steps of:
6-1, selecting key pedestrians
Processing short tracks of the pedestrians in each section of the video image i obtained in the step 5-5 in the sliding time window, and selecting key pedestrians in the short tracks in each section of the video image i by using the appearance similarity of the pedestrians for all the short tracks in one sliding time window, and using the key pedestrians to represent one short track;
step 6-2, combining every two key pedestrians in short tracks
In a sliding time window, combining every two key pedestrians with each short track in a certain section of video image i selected in the step 6-1 and every two key pedestrians with each short track in other sections of video images j, wherein each combination is two pictures with the size of 128 × 64 × 3 and containing the key pedestrians;
step 6-3, calculating the similarity of the short tracks
Inputting two pictures with the size of 128 × 64 × 3 containing key pedestrians into the twin neural network trained in the step 4, and outputting the result as the appearance similarity of the two key pedestrians with short tracksThen, the operation is carried out according to the following formula, and the similarity of the two short tracks is output:
wherein the content of the first and second substances,representing the appearance similarity of two short-track key pedestrians, and N representing the coincidence conditionK ofASSet, N represents the number of elements in N;
step 6-4, calculating eachSpeed of motion v of short track of striptracklet
The motion speed of each short track is calculated according to the following formula:
PSindicating the start of a short track, PeIndicating the end of a short track, TSIndicating the start of a short track, TeA termination frame representing a short track;
step 6-5, calculating the motion affinity Tr between the short tracksMA
TrMA=max(1-β*e(Tr1,Tr2),0)
e(Tr1,Tr2)=eforward(Tr1,Tr2)+ebackward(Tr1,Tr2)
Wherein P is1And P2For two short tracks, beta is the parameter to be adjusted, eXforward,eYforward,eXbackwardAnd eYbackwardIs the forward and reverse error of the short track in the X and Y directions, eforwardAnd ebackwardAre the forward error and the reverse error of the short track motion,andis the end point coordinate of a short track,andis the starting point coordinate of another short track, t1And t2V is obtained from the frame index of the end point coordinate of one short track and the frame index of the start point coordinate of the other short track, S0604trackletDecomposed into X-and Y-direction movement velocities, v1x,v1y,v2xAnd v2yThe moving speeds of the two short tracks in the X direction and the Y direction are respectively;
step 6-6, establishing a short track correlation matrix
Similarity Tr of short tracks obtained in step 6-3 by using sigmoid functionsAnd the kinematic affinity Tr of the short trajectory obtained in step 6-5MAAnd (3) fusing into a short track correlation matrix TrC:
where λ and μ are parameters for regulation;
step 6-7, solving classification
Solving and classifying the short track correlation matrix TrC obtained in the step 6-6 by using a maximum correlation principle, wherein each obtained class respectively represents the short tracks with the same identity, the maximum correlation principle is as follows, and a graph G ═ V, E, W is set, wherein a vertex V represents each short track, E represents an edge connected with the short tracks, and W represents the correlation between the two short tracks:
wherein v is1,v2,v3Representing short tracks in V, w represents the correlation degree between the short tracks, and the value of x is 0 or 1;
step 6-8, merging the growth tracks
Merging the short tracks of each type obtained by solving and classifying in the step 6-7, so as to merge the short tracks of the pedestrians with the same identity into a new track; as the sliding time window progresses, a complete long trajectory for all pedestrians is finally obtained.
4. A twin neural network based pedestrian tracking method according to claim 3, further characterized by: the step 4 of establishing and training the twin neural network comprises the following steps:
step 4-1, creating a twin neural network
A twin neural network is built by using a Pythrch frame, and the structure is as follows: two 128 × 64 × 3 pedestrian pictures are stacked into 128 × 64 × 6 and then used as the input of the twin neural network, and pass through three layers of convolution layers, wherein the convolution kernel size of the first layer is 9 × 12, the convolution kernel size of the second layer is 5 × 16, and the convolution kernel size of the third layer is 5 × 24; each layer convolution contains batch normalization, relu activation function, pooling of 2 x 2; the output of the three-layer convolution is subjected to one-dimensional operation to obtain a vector of 1 x 1152, then the vector passes through two full-connected layers with the sizes of 1 x 150 and 1 x 2, and finally the output result is the appearance similarity of two pedestrian pictures;
step 4-2, training twin neural network
Data set setting: downloading a CUHK02 pedestrian tracking data set from the Internet, wherein the data set comprises 1816 pedestrian identities, each pedestrian is shot by 2 camera angles, each camera angle has 2 photos, the first 80% of pedestrians are taken as a training set, and the rest pedestrians are taken as a testing set;
training process: two rounds of twin neural network training established in the step 4-1 are carried out, in the first round, training data are pictures of a first shooting angle of pedestrians in a training set, sample combinations are randomly generated in the training process, and the proportion of positive samples to negative samples is controlled to be 1: 1, the positive samples are two pedestrians with the same identity, and the negative samples are two pedestrians with different identities; in the second round, the picture of the second camera angle of the pedestrian in the training set is used as training data, and the generation mode of the positive sample and the negative sample is the same as that of the first round;
training and testing: after training is finished, testing the trained twin neural network by using the test set, wherein the output of the twin neural network is the appearance similarity of two pedestrians, and when the appearance similarity of the two pedestrians with the same identity is more than or equal to 0.5, the two pedestrians are considered to be correct; on the contrary, the two pedestrians with different identities are considered to be correct when the appearance similarity is less than or equal to 0.5; and finally, the accuracy of the test result reaches more than 98%, and the trained twin neural network is stored.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010584083.9A CN111814604A (en) | 2020-06-23 | 2020-06-23 | Pedestrian tracking method based on twin neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010584083.9A CN111814604A (en) | 2020-06-23 | 2020-06-23 | Pedestrian tracking method based on twin neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111814604A true CN111814604A (en) | 2020-10-23 |
Family
ID=72845520
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010584083.9A Pending CN111814604A (en) | 2020-06-23 | 2020-06-23 | Pedestrian tracking method based on twin neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111814604A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114845119A (en) * | 2022-07-04 | 2022-08-02 | 光谷技术有限公司 | Thing allies oneself with gateway and verifies and compression system |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150286873A1 (en) * | 2014-04-03 | 2015-10-08 | Bruce L. Davis | Smartphone-based methods and systems |
US20190147602A1 (en) * | 2017-11-13 | 2019-05-16 | Qualcomm Technologies, Inc. | Hybrid and self-aware long-term object tracking |
CN109784155A (en) * | 2018-12-10 | 2019-05-21 | 西安电子科技大学 | Visual target tracking method, intelligent robot based on verifying and mechanism for correcting errors |
CN110135314A (en) * | 2019-05-07 | 2019-08-16 | 电子科技大学 | A kind of multi-object tracking method based on depth Trajectory prediction |
CN110298404A (en) * | 2019-07-02 | 2019-10-01 | 西南交通大学 | A kind of method for tracking target based on triple twin Hash e-learnings |
AU2019101133A4 (en) * | 2019-09-30 | 2019-10-31 | Bo, Yaxin MISS | Fast vehicle detection using augmented dataset based on RetinaNet |
CN110619655A (en) * | 2019-08-23 | 2019-12-27 | 深圳大学 | Target tracking method and device integrating optical flow information and Simese framework |
CN110942471A (en) * | 2019-10-30 | 2020-03-31 | 电子科技大学 | Long-term target tracking method based on space-time constraint |
KR102095685B1 (en) * | 2019-12-02 | 2020-04-01 | 주식회사 넥스파시스템 | vehicle detection method and device |
-
2020
- 2020-06-23 CN CN202010584083.9A patent/CN111814604A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150286873A1 (en) * | 2014-04-03 | 2015-10-08 | Bruce L. Davis | Smartphone-based methods and systems |
US20190147602A1 (en) * | 2017-11-13 | 2019-05-16 | Qualcomm Technologies, Inc. | Hybrid and self-aware long-term object tracking |
CN109784155A (en) * | 2018-12-10 | 2019-05-21 | 西安电子科技大学 | Visual target tracking method, intelligent robot based on verifying and mechanism for correcting errors |
CN110135314A (en) * | 2019-05-07 | 2019-08-16 | 电子科技大学 | A kind of multi-object tracking method based on depth Trajectory prediction |
CN110298404A (en) * | 2019-07-02 | 2019-10-01 | 西南交通大学 | A kind of method for tracking target based on triple twin Hash e-learnings |
CN110619655A (en) * | 2019-08-23 | 2019-12-27 | 深圳大学 | Target tracking method and device integrating optical flow information and Simese framework |
AU2019101133A4 (en) * | 2019-09-30 | 2019-10-31 | Bo, Yaxin MISS | Fast vehicle detection using augmented dataset based on RetinaNet |
CN110942471A (en) * | 2019-10-30 | 2020-03-31 | 电子科技大学 | Long-term target tracking method based on space-time constraint |
KR102095685B1 (en) * | 2019-12-02 | 2020-04-01 | 주식회사 넥스파시스템 | vehicle detection method and device |
Non-Patent Citations (2)
Title |
---|
ZHANG HAN ET AL.: "Visual tracking using Siamese convolutional neural network with region proposal and domain specific updating", 《NEUROCOMPUTING》 * |
宗家辉: "基于TLD框架的目标跟踪算法的研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114845119A (en) * | 2022-07-04 | 2022-08-02 | 光谷技术有限公司 | Thing allies oneself with gateway and verifies and compression system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107832672B (en) | Pedestrian re-identification method for designing multi-loss function by utilizing attitude information | |
CN106709449B (en) | Pedestrian re-identification method and system based on deep learning and reinforcement learning | |
Yang et al. | An online learned CRF model for multi-target tracking | |
US20200226421A1 (en) | Training and using a convolutional neural network for person re-identification | |
Yan et al. | To track or to detect? an ensemble framework for optimal selection | |
Li et al. | Tracking in low frame rate video: A cascade particle filter with discriminative observers of different life spans | |
CN111709311B (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
Chen et al. | Aggregate tracklet appearance features for multi-object tracking | |
CN110197502B (en) | Multi-target tracking method and system based on identity re-identification | |
Kaâniche et al. | Recognizing gestures by learning local motion signatures of HOG descriptors | |
CN110660082A (en) | Target tracking method based on graph convolution and trajectory convolution network learning | |
Gupta et al. | Nose, eyes and ears: Head pose estimation by locating facial keypoints | |
CN109191497A (en) | A kind of real-time online multi-object tracking method based on much information fusion | |
CN112989889B (en) | Gait recognition method based on gesture guidance | |
Wong et al. | Track everything: Limiting prior knowledge in online multi-object recognition | |
CN109325546A (en) | A kind of combination footwork feature at time footprint recognition method | |
CN114283355A (en) | Multi-target endangered animal tracking method based on small sample learning | |
Arantes et al. | Human gait recognition using extraction and fusion of global motion features | |
CN113436231B (en) | Pedestrian track generation method, device, equipment and storage medium | |
CN111814604A (en) | Pedestrian tracking method based on twin neural network | |
Mademlis et al. | Exploiting stereoscopic disparity for augmenting human activity recognition performance | |
Cheheb et al. | Investigating the use of autoencoders for gait-based person recognition | |
Hashmi et al. | GAIT analysis: 3D pose estimation and prediction in defence applications using pattern recognition | |
CN114038011A (en) | Method for detecting abnormal behaviors of human body in indoor scene | |
Zhang et al. | What makes for good multiple object trackers? |
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 |