CN107358620B - Hybrid system-based full-coverage pedestrian tracking method and device - Google Patents

Hybrid system-based full-coverage pedestrian tracking method and device Download PDF

Info

Publication number
CN107358620B
CN107358620B CN201610303631.XA CN201610303631A CN107358620B CN 107358620 B CN107358620 B CN 107358620B CN 201610303631 A CN201610303631 A CN 201610303631A CN 107358620 B CN107358620 B CN 107358620B
Authority
CN
China
Prior art keywords
pedestrian
tracking
model
pedestrians
identification
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201610303631.XA
Other languages
Chinese (zh)
Other versions
CN107358620A (en
Inventor
胡士强
张晓宇
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.)
Shanghai Jiaotong University
Original Assignee
Shanghai Jiaotong 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 Shanghai Jiaotong University filed Critical Shanghai Jiaotong University
Priority to CN201610303631.XA priority Critical patent/CN107358620B/en
Publication of CN107358620A publication Critical patent/CN107358620A/en
Application granted granted Critical
Publication of CN107358620B publication Critical patent/CN107358620B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)
  • Traffic Control Systems (AREA)

Abstract

A fully-shielded pedestrian tracking method based on a hybrid system is characterized in that a corresponding tracking model and an identification model are constructed according to the motion characteristics of pedestrians and the characteristics of the pedestrians needing to be learned in tracking, when the pedestrians are not shielded, the tracking model tracks the pedestrians and learns the characteristics of the pedestrians at the same time, and when the pedestrians are shielded, the pedestrians are sensed to be shielded through the sensing conditions of a model transfer rule and are switched into the identification model; the identification model continuously identifies the lost-to-follow pedestrian from the detection result by utilizing the learned pedestrian characteristics according to the identification condition of the model transfer rule; when the identification model identifies a pedestrian, the state of tracking the pedestrian is reset through the reset condition of the model transfer rule and the tracking model is switched to continue tracking, so that the tracking of the pedestrian under the condition of full shelter is realized; the invention has reasonable design, integrates the idea of utilizing a hybrid system for model driving and data driving algorithms, solves the practical problem and greatly improves the accuracy and the success rate of tracking.

Description

Hybrid system-based full-coverage pedestrian tracking method and device
Technical Field
The invention relates to a technology in the field of tracking of computer vision monitoring targets, in particular to a full-shading pedestrian tracking method and device based on a hybrid system.
Background
There are two main categories of traditional single pedestrian tracking: target representation and localization, filtering and data association. The pedestrian tracking algorithm commonly used at present is the first type of method. Such methods can be further divided into three subclasses: generative tracking, discriminant tracking, and detection and tracking hybrid tracking. The generative algorithm learns that a representative model represents a tracking target and searches for an image region that minimizes reconstruction errors using the model; the discriminant algorithm regards the tracking problem as a two-classification problem in a local search region, and searches for an image region with the highest discrimination with the background in the current frame by using the learned classifier.
An algorithm based on target Representation and localization has been developed sufficiently in the last decade.A practical and effective algorithm is proposed for pedestrian occlusion, such as L1 tracking algorithm (Xue Mei and Haibin L ing, Robust visual tracking and Vehicle Classification view field tracking [ J ] reconstruction [ J ] IEEE TPAMI, November 2011, vol.33(11):2259-2271) converts tracking into a Sparse approximation problem under a particle filtering framework.during tracking, candidate targets are represented linearly by target templates and trivial templates.
A representative algorithm for the combination of detection and tracking is the T L D (ZDenk Kalal, Krystian Mikolajczyk, Jiri Matas. tracking-learning-detection [ J ]. IEEE T PAMI, July 2012,34(7): 1409) -.
Hybrid systems are a class of systems that contain both continuous and discrete dynamic behavior, which not only co-exist but also interact, the evolution of hybrid systems depends on both the response to discrete transient events and the response to time-varying dynamic behavior represented by differential and difference equations.
Due to the fact that the visual angle of a single camera in a monitoring scene is limited, the situation of full pedestrian shielding inevitably exists, and in the situation, the algorithm is required to recognize the shielding for tracking the pedestrian and continue to track when the pedestrian with the shielding appears again.
The search of the prior art shows that Chinese patent document No. CN102663409A, published Japanese 2012.9.12 discloses a pedestrian tracking method based on HOG-L BP description, which comprises the following steps of A1 sample establishment, A2 feature extraction, A3 SVM model establishment, A4 classifier training, A5 video capture and preprocessing, A6 video pedestrian detection, A7 video pedestrian tracking, wherein the step of tracking the pedestrian detected in the step A6 by using a particle filtering tracking method based on HOG-L BP features is carried out, but the technology 1) is not favorable for the detection of unknown pedestrian because a training sample for tracking the pedestrian is required to be added when the positive and negative samples are selected by using the HOG-L BP descriptor and the pedestrian detection algorithm of the SVM, 2) the condition that the pedestrian falls outside a particle filtering prediction region is not considered when the shielding problem is solved, 3) the algorithm has no dead track ending mechanism, the pedestrian is still updated after the pedestrian is shielded, the template is inevitably updated, the subsequent template is inevitably updated, and the pedestrian tracking algorithm is carried out if the pedestrian is not capable of being completely identified as a new generation target after the filtering prediction of the pedestrian is not described in the newly-predicted position.
Disclosure of Invention
The invention provides a full-shading pedestrian tracking method and device based on a hybrid system, aiming at the defects in the prior art, the pedestrian tracking model and the pedestrian identification model are constructed, the pedestrian characteristics are tracked and learned through the pedestrian tracking model before the pedestrian is fully shaded, the learning is stopped when the pedestrian is tracked and fully shaded, and the pedestrian with no follow is identified from the detection result through the pedestrian identification model by utilizing the learned pedestrian characteristics and is continuously tracked.
The invention is realized by the following technical scheme:
the invention relates to a full-shielding pedestrian tracking method based on a hybrid system, which comprises the steps of constructing a corresponding tracking model and an identification model according to the motion characteristics of pedestrians and the characteristics of the pedestrians needing to be learned in tracking, wherein when the pedestrians are not shielded, the tracking model tracks the pedestrians and simultaneously learns the characteristics of the pedestrians, and when the pedestrians are shielded, the pedestrians are sensed to be shielded through the sensing conditions of a model transfer rule and are switched into the identification model; the identification model continuously identifies the lost-to-follow pedestrian from the detection result by utilizing the learned pedestrian characteristics; when the identification model recognizes that the pedestrian is lost, the state of tracking the pedestrian is reset according to the reset condition of the model transfer rule and the tracking model is switched to continue tracking, so that the tracking of the pedestrian under the condition of full shielding is realized.
The detection result is as follows: and detecting the pedestrian area by adopting a pedestrian detection algorithm ACF.
The pedestrian detection result is expressed as
Figure BDA0000985384030000031
Wherein:
Figure BDA0000985384030000032
the location and scale of the ith measurement with noise for time t.
The tracking model is expressed as
Figure BDA0000985384030000033
Wherein: x is a parameter of the motion state of the pedestrian,
Figure BDA0000985384030000034
for a particular pedestrian classifier to be trained,
Figure BDA00009853840300000313
the average moving speed of the pedestrian is obtained,
Figure BDA00009853840300000314
to track average scale information of pedestrians.
The identification model comprises: a reachability sub-model, a scale sub-model and an appearance sub-model of the pedestrian movement.
The accessibility submodel is
Figure BDA0000985384030000035
Wherein: diAs a result of the detection
Figure BDA0000985384030000036
(1:2) center x of center and pedestrian tracked before tracking model failurek(1:2) in the case of
Figure BDA00009853840300000312
The reachability sub-model holds.
The scale sub model is P't|t-1＝Q′+F′P′t-1FT，S′＝H′P′t|t-1H′T+ R', wherein: z 'is scale information on the detection result containing noise, and F' ═ I2，H′＝I2，
Figure BDA0000985384030000037
If it is not
Figure BDA0000985384030000038
And if gamma' is the threshold of the two-dimensional elliptical goal, the scale sub-model is established.
The said appearance submodel is in
Figure BDA0000985384030000039
Figure BDA00009853840300000310
Is composed of
Figure BDA00009853840300000311
This is true for framed pedestrian image regions.
The invention relates to a device for realizing the method, which comprises the following steps: video acquisition unit, tracking and characteristic learning unit, identification element and transfer rule unit, wherein: the tracking and feature learning unit learns the features of the pedestrians while tracking the pedestrians, and the identification unit identifies the persons who are missed to follow from the detection result according to the features of the pedestrians learned by the tracking and feature learning unit; the transition rule unit performs switching and state resetting between the tracking and feature learning unit and the recognition unit according to the sensing condition, the recognition condition and the reset condition.
Technical effects
Compared with the prior art, the pedestrian tracking method is reasonable in design, easy to control and capable of greatly improving accuracy and success rate of pedestrian tracking by tracking pedestrians based on a hybrid system and realizing detection of unknown pedestrians by combining advanced pedestrian detection algorithm ACF with INRIA database without training sample information of the tracked pedestrians.
Drawings
FIG. 1 is a schematic diagram of a full-occlusion pedestrian tracking method;
FIG. 2 is a graph of qualitative results of the examples;
fig. 3 is a schematic diagram of a full-occlusion pedestrian tracking device.
Detailed Description
The following examples are given for the detailed implementation and specific operation of the present invention, but the scope of the present invention is not limited to the following examples.
Example 1
As shown in fig. 1, the present embodiment relates to a full-occlusion pedestrian tracking method based on a hybrid system, which includes the following steps:
step 1, according to the pedestrian areas of N pedestrians detected by the ACF, the pedestrian detection algorithm is used as a detection result, a tracking model is constructed by combining a Kalman filtering algorithm on the assumption that the pedestrian movement generally accords with a constant-speed movement model, the pedestrian tracking is realized, and the pedestrian characteristics are learned.
The pedestrian detection algorithm ACF (aggregated Channel features) is introduced in Fast features pyramids for object detection (J, IEEET PAMI, August 2014,36(8):1532-1545) of Piotr Dollar et al.
The pedestrian detection result is expressed as
Figure BDA0000985384030000042
Wherein:
Figure BDA0000985384030000043
representing the position and scale of the noisy ith measurement at time t.
The pedestrian tracking model
Figure BDA00009853840300000415
State at a certain moment
Figure BDA0000985384030000044
Wherein: x is a parameter of the motion state of the pedestrian,
Figure BDA0000985384030000045
in order to provide a pedestrian classifier that needs to be trained,
Figure BDA0000985384030000046
the average moving speed of the pedestrian is obtained,
Figure BDA0000985384030000047
is an average scale for tracking pedestrians.
The pedestrian motion state parameter
Figure BDA0000985384030000048
Wherein: (p)w,ph) Which is indicative of the position of the pedestrian,
Figure BDA0000985384030000049
representing the speed of the pedestrian, and (w, h) representing the width and height of the pedestrian.
The motion model and the measurement model of the pedestrian motion state parameter x are respectively as follows:
Figure BDA00009853840300000414
Figure BDA00009853840300000410
wherein:
Figure BDA00009853840300000411
and representing a Gaussian density function with mean and variance of m and P respectively, wherein F is a state transition matrix, Q is a process noise covariance matrix, H is an observation matrix, and R is a process noise covariance.
Since the motion of a pedestrian tends to follow a constant velocity motion, therefore:
Figure BDA00009853840300000412
Figure BDA00009853840300000413
wherein: sigmauIs the process noise standard deviation, σvTo observe the standard deviation of the noise.
The pedestrian classifier
Figure BDA00009853840300000416
The method specifically comprises the following steps: extracting N near the tracking result while tracking the pedestrianposA positive sample and NnegA negative sample and calculating a corresponding value according to the templateAnd training a classifier which can identify and track the pedestrian from the detection result based on the pedestrian characteristic by adopting an incremental SVM based on the values.
The positive sample is extracted in a circular area taking the center of the estimated pedestrian as the center and the Sw for tracking the width of the pedestrian as the radius, and the negative sample is extracted in a circular area taking the center of the estimated pedestrian and respectively taking the Swi and the Swo for tracking the width of the pedestrian as the inner radius and the outer radius.
The values of the positive sample and the negative sample are expressed by the similarity sim (T, y) of the extracted positive sample and the extracted negative sample and the template, wherein: t is the template image and y is the image of the positive and negative examples.
After the template image T and the images y of the positive sample and the negative sample are adjusted to the same image size (e.g. 128 × 64), a pedestrian similarity measurement algorithm or the like is used to perform pedestrian similarity measurement.
The pedestrian similarity measurement algorithm is introduced In Zhao et al, Unsurerved science L earning for Person Re-identification (In processing of conference Computer Vision and Pattern Recognition [ C ], Portland, OR, USA,25-27June 2013: 3586-3593).
The incremental SVM may be referred to as "SVM innovative learning, adaptive optimization" by Diehl et al (In Proceedings of the International Joint Conference on neural Networks [ C ], Portland, OR, USA,20-24July 2003: 2685-.
The average moving speed of the pedestrian
Figure BDA0000985384030000057
Wherein v is the current moving speed of the pedestrian, α is the learning parameter of the average speed of the pedestrian, and 0 is more than or equal to α is less than or equal to 1.
The average scale for tracking pedestrians
Figure BDA0000985384030000056
Wherein s is the current scale of the pedestrian, β is the learning parameter of the average scale of the pedestrian, and 0 is more than or equal to β is more than or equal to 1.
And 2, constructing an identification model, and screening a target which can be used for tracking the pedestrian from the detection result by combining the accessibility, the scale information and the apparent information of the pedestrian movement.
The identification model
Figure BDA0000985384030000058
The method comprises the following steps: a reachability sub-model, a scale sub-model and an appearance sub-model of the pedestrian movement.
The accessibility submodel is
Figure BDA0000985384030000055
Wherein: diAs a result of the detection
Figure BDA0000985384030000054
Tracking pedestrian center x before center and pedestrian tracking model failurek(1:2) in the case of
Figure BDA0000985384030000053
The reachability sub-model holds.
The scale sub model is P't|t-1＝Q′+F′P′t-1FT，S′＝H′P′t|t-1H′T+ R', wherein: z 'is scale information on the detection result containing noise, and F' ═ I2，H′＝I2，
Figure BDA0000985384030000051
If it is not
Figure BDA0000985384030000052
And if gamma' is the threshold of the two-dimensional elliptical goal, the scale sub-model is established.
The scale sub-model may employ an ellipsoid gate algorithm, as described in Design and analysis of model tracking systems by Blackman et al (Aretech House, Boston, 1999).
The two-dimensional ellipse goal threshold is introduced from the book of maneuvering target tracking (B national defense industry publishing, 1991) of Zhou-hongren et al.
The said appearance submodel is in
Figure BDA00009853840300000613
This is true. Wherein
Figure BDA00009853840300000614
The ith pedestrian at the time t
Figure BDA00009853840300000615
The information of the image to be framed,
Figure BDA00009853840300000612
is a particular pedestrian classifier learned in the tracking model.
And 3, establishing a model transfer rule, judging the model which acts at the current moment in real time, and carrying out experimental detection.
The model transition rule comprises a perception condition tau1Identification condition tau2And reset condition τ3。
The sensing condition tau1Comprises the following steps: is provided with
Figure BDA0000985384030000063
And there is a unique detection result satisfying Pt|t-1＝Q+FPt-1FT，
Figure BDA0000985384030000062
(gamma is the threshold of the four-dimensional elliptical goal), then
Figure BDA00009853840300000610
Xt＝Xt-1。
The identification condition tau2Comprises the following steps: is provided with
Figure BDA0000985384030000065
If pedestrian discerns the model
Figure BDA00009853840300000611
If not, then
Figure BDA0000985384030000064
Xt＝Xt-1。
The mode reset condition tau3Comprises the following steps: is provided with
Figure BDA0000985384030000066
If pedestrian discerns the model
Figure BDA0000985384030000067
Is established, then
Figure BDA0000985384030000068
Figure BDA0000985384030000069
The data detected by the experiment adopt a self-shot video with the pedestrian fully sheltered.
The video has great change in the posture before and after tracking the people to enter and exit the group.
After the tracked pedestrians are completely shielded, the video disturbs the first-out of the pedestrians and the position of the missed-out pedestrians is unknown.
The values of the parameters of this embodiment are shown in table 1.
Table 1 parameter setting table
Figure BDA0000985384030000061
As shown in fig. 2, which is a result of a qualitative experiment in the present embodiment, it can be known that the present embodiment can ensure that the tracked pedestrian loses the stable tracking before following; the blocking can be judged in time after the tracked pedestrian is blocked; the method comprises the following steps of not tracking when an interfering pedestrian appears, and identifying and tracking the pedestrian which is out of tracking in time after a real target appears; the algorithm is not affected by the reappearance position of the lost-to-follow pedestrian (the algorithm can identify and retrack the lost-to-follow pedestrian after the pedestrian is shielded, no matter whether the pedestrian walks in the forward direction of shielding or returns according to the original route).
The average center error of this example compared to FCT, L IAPG, and T L D algorithms in five test video groups is shown in table 2.
TABLE 2 comparison of mean center errors
Test1 Test2 Test3 Test4
FCT 28 40 64 17
TLD 45 49 26 13
L1APG 20 44 41 7
Ours 2 3 2 4
As can be seen from Table 2, since the pedestrian with missing heel can be accurately captured from the detection result after being completely shielded, the average center error of the embodiment in each set of test is the smallest. Other algorithms have a high average center error due to search range or poor pedestrian recognition.
The tracking success rate results of this example and the FCT, L IAPG, and T L D algorithms in four sets of test videos are shown in table 3.
Table 3 tracking success rate results
Test1 Test2 Test3 Test4
FCT 0.63 0.60 0.01 0.57
TLD 0.36 0.52 0.70 0.74
L1APG 0.65 0.60 0.53 0.91
Ours 0.99 0.93 0.96 0.93
As can be seen from table 3, the present embodiment has the highest success rate of pedestrian tracking when there is full occlusion, and remains stable in the test, and can realize timely termination of tracking after the pedestrian disappears, and can accurately recognize the target that is lost and track from the detection result.
The model transfer rule of the embodiment adopts an ellipsoid gate algorithm, and when the detection result is not only one within an ellipsoid gate for tracking pedestrians, the model needs to be transferred from the tracking model to the identification model; when the identification model fails to recognize the pedestrian from the detection result, the model jumps to the identification model again; when the recognition model recognizes a tracking pedestrian from its detection result, the pedestrian tracking position state is reset and transferred from the recognition model to the tracking model.
As shown in fig. 3, the present embodiment relates to an apparatus for implementing the method, including: video acquisition unit, tracking and characteristic learning unit, identification element and transfer rule unit, wherein: the tracking unit learns the characteristics of the pedestrians while tracking the pedestrians, and the identification unit identifies the persons who are missed to follow from the detection result according to the characteristics of the pedestrians learned by the tracking and characteristic learning unit; the transition rule unit performs switching and state resetting between the tracking and feature learning unit and the recognition unit according to the sensing condition, the recognition condition and the reset condition.

Claims (4)

1. A fully-shielded pedestrian tracking method based on a hybrid system is characterized in that a corresponding tracking model and an identification model are constructed according to the motion characteristics of pedestrians and the characteristics of the pedestrians needing to be learned in tracking, when the pedestrians are not shielded, the tracking model tracks the pedestrians and learns the characteristics of the pedestrians, and when the pedestrians are shielded, the pedestrians are sensed to be shielded through the sensing conditions of a model transfer rule and are switched into the identification model; the identification model continuously identifies the lost-to-follow pedestrian from the detection result by utilizing the learned pedestrian characteristics; when the identification model identifies a pedestrian, the state of tracking the pedestrian is reset through the reset condition of the model transfer rule and the tracking model is switched to continue tracking, so that the tracking of the pedestrian under the condition of full shelter is realized;
the detection result is as follows: a pedestrian region detected by adopting a pedestrian detection algorithm ACF;
the identification model comprises: a reachability sub-model, a scale sub-model and an appearance sub-model of pedestrian motion, wherein:
the accessibility submodel is
Figure FDA0002501915520000011
Wherein: diAs a result of the detection
Figure FDA0002501915520000012
Tracking pedestrian center x before center and pedestrian tracking model failurek(1:2) Euclidean distance of
Figure FDA0002501915520000013
The reachability sub-model is established and,
Figure FDA0002501915520000014
the average moving speed of the pedestrian is obtained;
the scale sub model is P't|t-1＝Q′+F′P′t-1FT，S′＝H′P′t|t-1H′T+ R', wherein: f is a state transition matrix, F ═ I2，H′＝I2，
Figure FDA0002501915520000015
When in use
Figure FDA0002501915520000016
If gamma' is the threshold of the two-dimensional elliptical goal, the scale sub-model is established;
the apparent submodel is in
Figure FDA0002501915520000017
The time is right; wherein
Figure FDA0002501915520000018
Is composed of
Figure FDA0002501915520000019
The pedestrian image area framed is the area of the pedestrian image,
Figure FDA00025019155200000110
is a particular pedestrian classifier learned in the tracking model.
2. The full obstruction pedestrian tracking method of claim 1, wherein said detection of a pedestrian is represented by
Figure FDA00025019155200000111
Wherein:
Figure FDA00025019155200000112
the position and scale of the ith detection result with noise at time t.
3. The full obstruction pedestrian tracking method of claim 1, wherein said tracking model is represented as
Figure FDA00025019155200000113
Wherein: x is a parameter of the motion state of the pedestrian,
Figure FDA00025019155200000114
to needThe pedestrian classifier to be trained is,
Figure FDA00025019155200000115
to track average scale information of pedestrians.
4. An apparatus for implementing the method of any of claims 1 to 3, comprising: video acquisition unit, tracking and characteristic learning unit, identification element and transfer rule unit, wherein: the tracking and feature learning unit learns the features of the pedestrians while tracking the pedestrians, and the identification unit establishes an identification model according to the features of the pedestrians learned by the feature learning unit and identifies the pedestrian with missing tracking in the detection result; the transfer rule unit performs switching and state resetting between the tracking unit and the identifying unit according to the sensing condition, the identifying condition and the resetting condition.
CN201610303631.XA 2016-05-10 2016-05-10 Hybrid system-based full-coverage pedestrian tracking method and device Active CN107358620B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610303631.XA CN107358620B (en) 2016-05-10 2016-05-10 Hybrid system-based full-coverage pedestrian tracking method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610303631.XA CN107358620B (en) 2016-05-10 2016-05-10 Hybrid system-based full-coverage pedestrian tracking method and device

Publications (2)

Publication Number Publication Date
CN107358620A CN107358620A (en) 2017-11-17
CN107358620B true CN107358620B (en) 2020-07-10

Family

ID=60271264

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610303631.XA Active CN107358620B (en) 2016-05-10 2016-05-10 Hybrid system-based full-coverage pedestrian tracking method and device

Country Status (1)

Country Link
CN (1) CN107358620B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2007033286A2 (en) * 2005-09-13 2007-03-22 Verificon Corporation System and method for object tracking and activity analysis
CN102646279A (en) * 2012-02-29 2012-08-22 北京航空航天大学 Anti-shielding tracking method based on moving prediction and multi-sub-block template matching combination
CN102881022A (en) * 2012-07-20 2013-01-16 西安电子科技大学 Concealed-target tracking method based on on-line learning
CN104951758A (en) * 2015-06-11 2015-09-30 大连理工大学 Vehicle-mounted method and vehicle-mounted system for detecting and tracking pedestrians based on vision under urban environment

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2007033286A2 (en) * 2005-09-13 2007-03-22 Verificon Corporation System and method for object tracking and activity analysis
CN102646279A (en) * 2012-02-29 2012-08-22 北京航空航天大学 Anti-shielding tracking method based on moving prediction and multi-sub-block template matching combination
CN102881022A (en) * 2012-07-20 2013-01-16 西安电子科技大学 Concealed-target tracking method based on on-line learning
CN104951758A (en) * 2015-06-11 2015-09-30 大连理工大学 Vehicle-mounted method and vehicle-mounted system for detecting and tracking pedestrians based on vision under urban environment

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
A cascade framework for unoccluded and occluded pedestrian detection;Aayush Ankit等;《Proceedings of the 2014 IEEE Students" Technology Symposium》;20140501;62-67 *

Also Published As

Publication number Publication date
CN107358620A (en) 2017-11-17

Similar Documents

Publication Publication Date Title
LaLonde et al. Clusternet: Detecting small objects in large scenes by exploiting spatio-temporal information
US11195038B2 (en) Device and a method for extracting dynamic information on a scene using a convolutional neural network
Choudhury et al. Vehicle detection and counting using haar feature-based classifier
Yahiaoui et al. Fisheyemodnet: Moving object detection on surround-view cameras for autonomous driving
Zhou et al. Robust vehicle detection in aerial images using bag-of-words and orientation aware scanning
Cao et al. Vehicle detection and motion analysis in low-altitude airborne video under urban environment
CN106778712B (en) Multi-target detection and tracking method
CN106886216B (en) Robot automatic tracking method and system based on RGBD face detection
CN108416780B (en) Object detection and matching method based on twin-region-of-interest pooling model
Zhang et al. Detection and tracking of multiple humans with extensive pose articulation
CN103246896A (en) Robust real-time vehicle detection and tracking method
Denman et al. Multi-spectral fusion for surveillance systems
Teutsch et al. Robust detection of moving vehicles in wide area motion imagery
Vignesh et al. Abnormal event detection on BMTT-PETS 2017 surveillance challenge
Su et al. A new local-main-gradient-orientation HOG and contour differences based algorithm for object classification
Amrouche et al. Vehicle Detection and Tracking in Real-time using YOLOv4-tiny
CN107292908A (en) Pedestrian tracting method based on KLT feature point tracking algorithms
Miller et al. Person tracking in UAV video
Zhang et al. An optical flow based moving objects detection algorithm for the UAV
Hu et al. Effective moving object detection from videos captured by a moving camera
CN110349184B (en) Multi-pedestrian tracking method based on iterative filtering and observation discrimination
Liu et al. Multi-view vehicle detection and tracking in crossroads
Yang et al. Edge-based moving object tracking algorithm for an embedded system
CN107358620B (en) Hybrid system-based full-coverage pedestrian tracking method and device
Ma et al. Depth assisted occlusion handling in video object tracking

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant