CN112560651A - Target tracking method and device based on combination of depth network and target segmentation - Google Patents

Target tracking method and device based on combination of depth network and target segmentation Download PDF

Info

Publication number
CN112560651A
CN112560651A CN202011449867.7A CN202011449867A CN112560651A CN 112560651 A CN112560651 A CN 112560651A CN 202011449867 A CN202011449867 A CN 202011449867A CN 112560651 A CN112560651 A CN 112560651A
Authority
CN
China
Prior art keywords
tracking
target
result
algorithm
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.)
Granted
Application number
CN202011449867.7A
Other languages
Chinese (zh)
Other versions
CN112560651B (en
Inventor
胡硕
杨莹光
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.)
Yanshan University
Original Assignee
Yanshan 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 Yanshan University filed Critical Yanshan University
Priority to CN202011449867.7A priority Critical patent/CN112560651B/en
Publication of CN112560651A publication Critical patent/CN112560651A/en
Application granted granted Critical
Publication of CN112560651B publication Critical patent/CN112560651B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/49—Segmenting video sequences, i.e. computational techniques such as parsing or cutting the sequence, low-level clustering or determining units such as shots or scenes
    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/215—Motion-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Geometry (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a target tracking method and a target tracking device based on the combination of a depth network and target segmentation, wherein the method comprises the following steps: acquiring a target image and a global image, and performing target tracking on the target area image and the global image by using a twin network algorithm to obtain a tracking result; carrying out target tracking on the global image by using a target detection algorithm to obtain a detection result; detecting the tracking drift condition of objects among the same type, and correcting the tracking result by using a correction strategy based on a target tracking track; preferentially selecting the corrected tracking result and the corrected detection result to obtain a primary coordinate position; and utilizing a segmentation algorithm to segment the initial coordinate position to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result. The method integrates the advantages of a twin network algorithm and a target segmentation algorithm, improves the tracking precision, ensures the running speed, and improves the inter-class tracking drift problem of the twin network.

Description

Target tracking method and device based on combination of depth network and target segmentation
Technical Field
The invention relates to the technical field of computer vision tracking, in particular to a target tracking method and device based on combination of a depth network and target segmentation.
Background
Artificial intelligence is rapidly developed, the comprehension capability of machines to various information is greatly improved, and computer vision is an important component in the field of artificial intelligence. The target tracking problem is a key problem in the field of machine vision, and the target tracking is more and more emphasized by more people due to wide application scenes. At present, the computing power of computers is increasing day by day, the further improvement of hardware facilities continuously pushes the development of the field of artificial intelligence, and the research of the target tracking problem is concerned more and more.
At present, the target tracking has wide practicability prospect and a plurality of application scenes, and is an important component of artificial intelligence. In daily life, the system comprises fire monitoring, traffic accident monitoring, race condition analysis of sports competitions, tracking function design of a security system and the like. However, the current target tracking field also has a plurality of problems and challenges, such as the occlusion problem during the tracking process, the low-resolution image blur problem, the complex background and the interference factor, and the like, and the above problems have high requirements on the tracking algorithm. In the current research of tracking algorithm, the accuracy and the running speed can be ensured at the same time, and the balancing of the two indexes is a main challenge.
Reviewing the development process of the target tracking algorithm, the traditional tracking algorithm in the past mainly realizes tracking according to target modeling or global search of target features. Two tracking concepts are represented here: a method of modeling based on an object model and a method based on a search. The model-based modeling method is to model the appearance model of the target and then find the target in the video frame information, such as region matching, feature point tracking, contour-based tracking algorithm, optical flow method, etc. In the above method, the most common method is a feature matching method, which extracts features of the target, and then finds out features with the highest similarity in subsequent frames to perform target positioning, where the most common features include SIFT features, SUFT features, Harris corner features, and the like. However, the disadvantage is also obvious, and people find that the method based on the target model modeling needs to process the whole image, and the real-time performance is poor due to large calculation amount. Based on a search method, people add a prediction algorithm into a tracking algorithm, and perform target search near the prediction value, so that the search range is reduced, the speed is further improved, and the search algorithm is commonly as follows: kalman filtering and particle filtering.
In summary, the conventional target tracking algorithms have a drawback that background information is not taken into account, so that poor performance is achieved under the condition of complex background, for example, tracking failure is very easy to occur under the conditions of illumination change, target occlusion and picture blurring. Moreover, the tracking speed can only reach about 10 frames per second, and the speed can not meet the real-time requirement.
Disclosure of Invention
The invention aims to provide a target tracking method and a target tracking device based on combination of a deep network and target segmentation, which combine a twin network algorithm and a target segmentation algorithm in deep learning to track a target and improve the precision and stability of target tracking on the basis of keeping the advantage of the twin network in the target tracking speed. In addition, the invention is also improved on the basis of the twin network, and the tracking drift problem among the congenital defects of the twin network is inhibited by utilizing a method based on tracking target track, thereby further improving the target tracking precision.
In order to achieve the above purpose, the invention provides the following technical scheme:
the invention provides a target tracking algorithm based on the combination of a depth network and target segmentation, which comprises the following steps:
step 1, obtaining a target image and a global image, and performing target tracking on the target area image and the global image by using a twin network algorithm to obtain a tracking result; carrying out target tracking on the global image by using a target detection algorithm to obtain a detection result;
step 2, detecting the tracking drift condition of objects among the same type, and correcting the tracking result by using a correction strategy based on a target tracking track;
step 3, carrying out preferential selection on the corrected tracking result and the corrected detection result to obtain a primary coordinate position;
and 4, segmenting the initial coordinate position by utilizing a segmentation algorithm to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result.
Further, correcting the tracking result by using a correction strategy based on the target tracking track, including:
under the condition that the number of similar objects exceeds a preset value, correcting the tracking result by utilizing a track-based scoring strategy;
under the condition that the number of the similar objects does not exceed a preset value, correcting the tracking result by utilizing a traversal correction strategy;
the traversal correction strategy comprises: when the algorithm detects that the track drifts, inquiring the positions of the objects with the highest scores of the current frame and arranging the objects in sequence, wherein the positions with the highest scores and close scores respectively correspond to the positions of similar objects of the same type; respectively traversing corresponding positions to carry out calculation, wherein the minimum calculated track drift is the corrected tracking result;
the trajectory-based scoring strategy includes: when the algorithm detects that the track drifts, a track-based correction function is added in the score calculation of the twin network algorithm to correct the score so as to inhibit the problem of tracking drift among the same classes; the method comprises the following steps of taking the track of the first five frames closest to a current frame, wherein the coordinate information is respectively as follows: l is1…L5The current frame coordinate is L6(ii) a The correction function comprises a first correction term D1And a second correction term D2Respectively as follows:
Figure BDA0002826356510000031
Figure BDA0002826356510000032
Figure BDA0002826356510000033
wherein D is1Represents the current frame L6And the previous frame L5Distance between, and average distance L of the previous five frames of the current frameAverageComparing, and correcting if the current frame has large position offset; d2Represents the current frame L6The first five frames and the last frame L5Variance comparison between the previous five frames if the current frame L6Correcting if the variance of the first five frames is large; omega1And ω2Weights of the first correction term and the second correction term respectively; score1And Score is the initial Score and the corrected Score calculated by the twin network, respectively; and s is a correction parameter which has a correction effect on the final value of the formula.
Further, the preferentially selecting the corrected tracking result and the detection result comprises the following steps:
the preferred selection is carried out by utilizing a preferred selection algorithm C (L)1,L2) The following were used:
Figure BDA0002826356510000041
wherein L is1To represent the boxes of the trace results, L2A group of boxes representing i pieces of position information in the detection result, wherein i is a positive integer; IoU, the coincidence ratio between the box representing the detection result and the box representing the tracking result is calculated as follows:
Figure BDA0002826356510000042
wherein the content of the first and second substances,
Figure BDA0002826356510000043
to representThe coincidence ratio between the box representing the tracking result and the box representing the ith positioning information in the detection result, and j is the value of i corresponding to IoU; the formula is as follows:
Figure BDA0002826356510000044
wherein the content of the first and second substances,
Figure BDA0002826356510000045
to represent the area of the box of the trace result,
Figure BDA0002826356510000046
is the area of the box representing the ith position information in the detection result.
Further, the segmenting the preliminary coordinate location using a segmentation algorithm includes: and (3) segmenting the initial coordinate position by adopting a Snake segmentation algorithm based on a depth network, wherein the final tracking result is expressed as follows:
L(x,y)＝Snake(L)；
L＝{(Max(xj),Max(yj))，(Min(xj),Min(yj))}；
wherein L is(x,y)The method comprises the following steps that a plurality of node coordinates of a segmented contour are obtained, L is a box representing a final tracking result, and the box takes the maximum value and the minimum value of a group of node coordinates in the plurality of node coordinates of the segmented contour as the upper left corner and the lower right corner; max (x)j) Is the maximum value of the abscissa, Max (y) in the coordinates of a plurality of nodes of the segmented contourj) Is the maximum value of the ordinate, Min (x), in the coordinates of a plurality of nodes of the segmented contourj) Is the minimum value of the abscissa, Min (y) in the coordinates of a plurality of nodes of the segmented contourj) The minimum value of the ordinate in the coordinates of a plurality of nodes of the segmented contour is obtained.
The invention also provides a high-precision target tracking device based on the depth network and the target segmentation, which comprises the following components:
the initial positioning module is used for acquiring a target image and a global image, and performing target tracking on the target area image and the global image by utilizing a twin network algorithm to obtain a tracking result; carrying out target tracking on the global image by using a target detection algorithm to obtain a detection result;
the tracking drift suppression module is used for detecting the tracking drift condition of objects among the same type and correcting the tracking result obtained by the initial positioning module by utilizing a target tracking track-based correction strategy;
the preferred selection module is used for carrying out preferred selection in the corrected tracking result and the detection result to obtain a preliminary coordinate position;
and the segmentation module is used for segmenting the initial coordinate position obtained by the optimization selection module by utilizing a segmentation algorithm to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result.
Further, the tracking drift suppression module is specifically configured to:
under the condition that the number of similar objects exceeds a preset value, correcting the tracking result by utilizing a track-based scoring strategy;
under the condition that the number of the similar objects does not exceed a preset value, correcting the tracking result by utilizing a traversal correction strategy;
the traversal correction strategy comprises: when the algorithm detects that the track drifts, inquiring the positions of the objects with the highest scores of the current frame and arranging the objects in sequence, wherein the positions with the highest scores and close scores respectively correspond to the positions of similar objects of the same type; respectively traversing corresponding positions to carry out calculation, wherein the minimum calculated track drift is the corrected tracking result;
the trajectory-based scoring strategy includes: when the algorithm detects that the track drifts, a track-based correction function is added in the score calculation of the twin network algorithm to correct the score so as to inhibit the problem of tracking drift among the same classes; the method comprises the following steps of taking the track of the first five frames closest to a current frame, wherein the coordinate information is respectively as follows: l is1…L5The current frame coordinate is L6(ii) a The correction function comprises a first correction term D1And a second correction term D2Respectively as follows:
Figure BDA0002826356510000061
Figure BDA0002826356510000062
Figure BDA0002826356510000063
wherein D is1Represents the current frame L6And the previous frame L5Distance between, and average distance L of the previous five frames of the current frameAverageComparing, and correcting if the current frame has large position offset; d2Represents the current frame L6The first five frames and the last frame L5Variance comparison between the previous five frames if the current frame L6Correcting if the variance of the first five frames is large; omega1And ω2Weights of the first correction term and the second correction term respectively; score1And Score is the initial Score and the corrected Score calculated by the twin network, respectively; and s is a correction parameter which has a correction effect on the final value of the formula.
Further, the preference selection module is specifically configured to:
the preferred selection is carried out by utilizing a preferred selection algorithm C (L)1,L2) The following were used:
Figure BDA0002826356510000064
wherein L is1To represent the boxes of the trace results, L2A group of boxes representing i pieces of position information in the detection result, wherein i is a positive integer; IoU, the coincidence ratio between the box representing the detection result and the box representing the tracking result is calculated as follows:
Figure BDA0002826356510000065
wherein the content of the first and second substances,
Figure BDA0002826356510000066
indicating the coincidence ratio between the box representing the tracking result and the box representing the ith positioning information in the detection result, wherein j is IoU corresponding to the value of i; the formula is as follows:
Figure BDA0002826356510000067
wherein the content of the first and second substances,
Figure BDA0002826356510000068
to represent the area of the box of the trace result,
Figure BDA0002826356510000069
is the area of the box representing the ith position information in the detection result.
Further, the segmentation module is specifically configured to:
and (3) segmenting the initial coordinate position by adopting a Snake segmentation algorithm based on a depth network, wherein the final tracking result is expressed as follows:
L(x,y)＝Snake(L)；
L＝{(Max(xj),Max(yj))，(Min(xj),Min(yj))}；
wherein L is(x,y)The method comprises the following steps that a plurality of node coordinates of a segmented contour are obtained, L is a box representing a final tracking result, and the box takes the maximum value and the minimum value of a group of node coordinates in the plurality of node coordinates of the segmented contour as the upper left corner and the lower right corner; max (x)j) Is the maximum value of the abscissa, Max (y) in the coordinates of a plurality of nodes of the segmented contourj) Is the maximum value of the ordinate, Min (x), in the coordinates of a plurality of nodes of the segmented contourj) Is the minimum value of the abscissa, Min (y) in the coordinates of a plurality of nodes of the segmented contourj) The minimum value of the ordinate in the coordinates of a plurality of nodes of the segmented contour is obtained.
The invention has the following beneficial effects:
(1) the image segmentation method is introduced into the target tracking, so that the target tracking precision is higher.
(2) And designing a combination scheme of the traditional twin network and an image segmentation algorithm, and providing a target detection result and tracking result preference strategy.
(3) And selecting an implementation scheme of an image segmentation algorithm, so that the speed of the whole scheme is ensured to reach real-time performance.
(4) And an inter-class drift scheme based on trajectory suppression of the twin network is provided, so that the congenital defect of the twin network is improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a flowchart of a target tracking method based on a combination of a depth network and target segmentation in an embodiment of the present invention;
FIG. 2 is a block diagram of a target tracking method based on a combination of a depth network and target segmentation in an embodiment of the present invention;
FIG. 3 is a flowchart of a target tracking method based on a combination of a depth network and target segmentation in an embodiment of the present invention;
FIG. 4 is a graph showing the results of the example of the present invention;
FIG. 5 is a graph showing still another result of the embodiment of the present invention.
Detailed Description
In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
Referring to fig. 1, a flowchart of a target tracking method based on a combination of a deep network and target segmentation in an embodiment of the present invention is shown, where the method includes:
step 1, obtaining a target image and a global image, and performing target tracking on the target area image and the global image by using a twin network algorithm to obtain a tracking result; carrying out target tracking on the global image by using a target detection algorithm to obtain a detection result;
step 2, detecting the tracking drift condition of objects among the same type, and correcting the tracking result by using a correction strategy based on a target tracking track;
the twin network algorithm is a classic target tracking algorithm, can well realize the real-time tracking effect, has obvious advantages in speed, but the output result of the twin network algorithm cannot reach excellent level in precision, and the traditional twin network adopts related operations, so that positioning drift is easy to occur in the tracking of similar objects with similarity. On the basis of a twin network algorithm, the target segmentation algorithm is introduced for fusion, and an inter-class drift suppression module is designed to obtain a stable and high-precision tracking effect.
Step 3, carrying out preferential selection on the corrected tracking result and the corrected detection result to obtain a primary coordinate position;
and 4, segmenting the initial coordinate position by utilizing a segmentation algorithm to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result.
According to the target tracking method based on the combination of the deep network and the target segmentation, the target tracking is performed by combining the twin network algorithm and the target segmentation algorithm in the deep learning, and the precision and the stability of the target tracking are improved on the basis of keeping the advantage of the twin network in the target tracking speed. In addition, the embodiment of the invention is improved on the basis of the twin network, and the tracking drift problem among the congenital defects of the twin network is inhibited by utilizing a method based on tracking the target track, so that the target tracking precision is further improved.
For the convenience of understanding, the following describes a target tracking method based on a combination of a deep network and target segmentation in the present invention with a specific example. Referring to fig. 2 and 3, fig. 2 shows a structural block diagram of a target tracking method based on a combination of a depth network and target segmentation in an embodiment of the present invention, and fig. 3 shows a flowchart of the target tracking method based on a combination of the depth network and target segmentation in the embodiment of the present invention. In this embodiment, an automobile is taken as a tracking target, and the method includes the following steps:
step 1, obtaining a target image and a global image, respectively tracking and detecting a target in the image by using a traditional twin network (Sim) algorithm and a target detection algorithm, and obtaining preliminary positioning information as a reference;
wherein, the step 1 is carried out according to the following two steps when in concrete implementation:
step 1.1, carrying out target tracking on the target area image and the global image by using a Sim algorithm to obtain a tracking result L1；
L1＝Siam(P1，P2) (1)
The traditional Sim algorithm combines the target image P with the target image P1And a global picture P2Respectively inputting the same network branches for processing, then performing related operation, and taking the position with the most obvious response as a first positioning position L1. But because the structure of the network is simpler and is influenced by the accuracy of the response position, the final accuracy is not good and only can be used as a reference, and therefore, the result of the Sim is used as a reference position by the invention.
Step 1.2, carrying out target tracking on the global image by using a target detection algorithm to obtain a detection result L2(i)(ii) a Wherein i is a positive integer;
objects in a scene are preliminarily classified by using an object classification algorithm, and the classification algorithm can accurately give position information of a specific classification object, for example, if the object is an automobile in the embodiment, the classification algorithm can accurately find all automobiles in the image.
L2(i)＝D(P2) (2)
Target classification algorithm D (P)2) All object classifications in the global image can be detected, and a plurality of example object positions detected in the global image, namely detection results, are obtained. In the embodiment of the invention, a mature target classification algorithm such as an RPN algorithm is adopted, and the obtained classification position information is accurate.
And 2, detecting the tracking drift condition of the objects among the same type, and correcting the tracking result by using a correction strategy based on a target tracking track.
The traditional twin network tracking algorithm causes tracking drift because the correlation scores of similar objects between two classes at a close distance are close by using correlation operation. In response to this problem, the present invention provides a correction strategy based on the target tracking trajectory to suppress inter-class drift.
According to the two conditions that the number of similar objects between classes is small and large, the invention respectively designs a traversal correction strategy and a score correction strategy based on the track.
Aiming at the condition that the number of similar objects does not exceed a preset value, a traversal correction strategy is designed, and the specific strategy is as follows:
when the algorithm detects that the track drifts, the algorithm queries the current frames, the scores of the current frames are arranged in sequence, and the positions with the highest scores and close scores respectively correspond to the positions of similar objects of the same type. And respectively carrying out traversal and substitution calculation on the corresponding positions, wherein the calculated trace with the minimum drift is the real position of the tracking target, namely the corrected tracking result.
Aiming at the condition that the number of the similar objects exceeds a preset value, the calculation amount of the traversal correction strategy is increased in multiples, so that a score correction strategy based on the track is designed; the specific strategy content is as follows: a track-based correction term is added in the calculation of the related operation scores of the traditional twin network, and the related operation scores of the twin network are corrected when the targets drift among classes, so that the problem of inter-class tracking drift is suppressed. The specific implementation can be carried out according to the following three steps:
the track of the first five frames closest to the current frame is taken, and the coordinate information is respectively as follows: l is1…L5The current frame coordinate is L6。
Step 2.1, design the first correction term D1The correction term represents the current frame position L6And last frame position L5The distance therebetween and the average distance LAverageIf the current frame position L6And last frame position L5Is much larger than the average distance LAverageThen, it is determined that tracking drift occurs.
Figure BDA0002826356510000111
Step 2.2, designing a second correction term D2The correction term represents the first five frames L of the previous frame1…L5Coordinate stability and the first five frames L of the current frame2…L6The coordinate stability of each other. If the current frame L6If the variance becomes large, it can be determined that tracking drift occurs.
Figure BDA0002826356510000112
Step 2.3, Score1And Score is the initial Score and the revised Score of the twin network, respectively;
Figure BDA0002826356510000113
wherein, ω is1And ω2The weights of the first correction term and the second correction term are respectively, and s is a correction parameter and plays a correction role in the final value of the formula.
And 3, performing preferred selection in the corrected tracking result and the detection result to obtain a preliminary coordinate position.
The tracking result is close to the target position but the positioning accuracy is not high, the detection result can provide positioning information of the classes of a plurality of groups of targets, the accuracy is high, and the information of which group is the target cannot be judged. For example, as shown in fig. 5, the target is a white car in the center of the image, fig. 4(b) is the tracking result, and fig. 4(a) is the detection result, which gives the location of the "car" category in several sets of images.
The preferred selection is carried out by utilizing a preferred selection algorithm C (L)1,L2) The following were used:
Figure BDA0002826356510000121
wherein L is1To represent the boxes of the trace results, L2A group of boxes representing i pieces of position information in the detection result, wherein i is a positive integer; IoU, the coincidence ratio between the box representing the detection result and the box representing the tracking result is calculated as follows:
Figure BDA0002826356510000122
wherein the content of the first and second substances,
Figure BDA0002826356510000123
indicating the coincidence ratio between the box representing the tracking result and the box representing the ith positioning information in the detection result, wherein j is IoU corresponding to the value of i; the formula is as follows:
Figure BDA0002826356510000124
wherein the content of the first and second substances,
Figure BDA0002826356510000125
to represent the area of the box of the trace result,
Figure BDA0002826356510000126
is the area of the box representing the ith position information in the detection result.
Equation (6) is a preferred selection algorithm, for L1，L2The represented position information is calculated IoU respectively and compared based on the maximum value, IoU represents the coincidence proportion between the detection result box and the tracking result box. If IoU is greater than 0.6, the target position tracked by the Sim is greatly overlapped with the target detection result position, and the target detection result position L is used2If the result is positive, otherwise, the result shows that the target detection algorithm does not detect the target, and L is used1The standard is. As shown in FIG. 5, the outermost and second outer borders represent L, respectively1，L2The algorithm selects more accurate L2The location information provides a reference for subsequent segmentation.
And 4, segmenting the initial coordinate position by utilizing a segmentation algorithm to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result.
In specific implementation, based on the preliminary coordinate position L obtained in the step 3, a depth-based Snake algorithm is adopted to segment the target.
L(x,y)＝Snake(L) (9)
L＝{(Max(xj),Max(yj))，(Min(xj),Min(yj))}； (10)
Wherein L is(x,y)The method comprises the following steps that a plurality of node coordinates of a segmented contour are obtained, L is a box representing a final tracking result, and the box takes the maximum value and the minimum value of a group of node coordinates in the plurality of node coordinates of the segmented contour as the upper left corner and the lower right corner; max (x)j) Is the maximum value of the abscissa, Max (y) in the coordinates of a plurality of nodes of the segmented contourj) Is the maximum value of the ordinate, Min (x), in the coordinates of a plurality of nodes of the segmented contourj) Is the minimum value of the abscissa, Min (y) in the coordinates of a plurality of nodes of the segmented contourj) The minimum value of the ordinate in the coordinates of a plurality of nodes of the segmented contour is obtained.
The Snake algorithm based on depth is used for segmenting the target, has the advantage of high speed, can achieve a real-time effect, and finally obtains a plurality of contour node positions L surrounding the object(x,y)And connecting a plurality of contour nodes in sequence to obtain the contour. For example, in fig. 5, the innermost ring-shaped contour of the target white car is the final segmentation result, i.e., the final algorithm result. In specific implementation, the extreme points of the abscissa and the ordinate in the contour nodes are taken to draw a rectangular contour.
As shown in fig. 4 and 5, which are graphs respectively showing the results obtained by the embodiment of the present invention. As shown in fig. 4, the tracking result obtained by the twin network algorithm for preliminary tracking and the detection result obtained by the target detection algorithm are shown, where the tracking result includes a position information, the detection results obtained by the target detection algorithm include a plurality of detection results, the white car in the image is used as the tracking target, the tracking result includes only one car, and the detection result includes a plurality of cars. As shown in fig. 5, the position with the largest IoU value is preferentially selected by the preferred selection algorithm, and the segmentation algorithm is operated based on the position. The result shows that the precision of the invention is obviously improved on the basis of the traditional Sim algorithm.
Corresponding to the target tracking method based on the combination of the depth network and the target segmentation in the invention, the invention also provides a target tracking device based on the combination of the depth network and the target segmentation, which comprises the following steps:
the initial positioning module is used for acquiring a target image and a global image, and performing target tracking on the target area image and the global image by utilizing a twin network algorithm to obtain a tracking result; carrying out target tracking on the global image by using a target detection algorithm to obtain a detection result;
the tracking drift suppression module is used for detecting the tracking drift condition of objects among the same type and correcting the tracking result obtained by the initial positioning module by utilizing a target tracking track-based correction strategy;
the preferred selection module is used for carrying out preferred selection in the corrected tracking result and the detection result to obtain a preliminary coordinate position;
and the segmentation module is used for segmenting the initial coordinate position obtained by the optimization selection module by utilizing a segmentation algorithm to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result.
Further, the tracking drift suppression module is specifically configured to:
under the condition that the number of similar objects exceeds a preset value, correcting the tracking result by utilizing a track-based scoring strategy;
under the condition that the number of the similar objects does not exceed a preset value, correcting the tracking result by utilizing a traversal correction strategy;
the traversal correction strategy comprises: when the algorithm detects that the track drifts, inquiring the positions of the objects with the highest scores of the current frame and arranging the objects in sequence, wherein the positions with the highest scores and close scores respectively correspond to the positions of similar objects of the same type; respectively traversing corresponding positions to carry out calculation, wherein the minimum calculated track drift is the corrected tracking result;
the trajectory-based scoring strategy includes: when the algorithm detects that the track drifts, a track-based correction function is added in the score calculation of the twin network algorithm to correct the score so as to inhibit the problem of tracking drift among the same classes; the method comprises the following steps of taking the track of the first five frames closest to a current frame, wherein the coordinate information is respectively as follows: l is1…L5The current frame coordinate is L6(ii) a The correction function comprises a first correction term D1And a second correction term D2Respectively as follows:
Figure BDA0002826356510000141
Figure BDA0002826356510000142
Figure BDA0002826356510000151
wherein D is1Represents the current frame L6And the previous frame L5Distance between, and average distance L of the previous five frames of the current frameAverageComparing, and correcting if the current frame has large position offset; d2Represents the current frame L6The first five frames and the last frame L5Variance comparison between the previous five frames if the current frame L6Correcting if the variance of the first five frames is large; omega1And ω2Weights of the first correction term and the second correction term respectively; score1And Score is the initial Score and the corrected Score calculated by the twin network, respectively; and s is a correction parameter which has a correction effect on the final value of the formula.
Further, the preference selection module is specifically configured to:
the preferred selection is carried out by utilizing a preferred selection algorithm C (L)1,L2) The following were used:
Figure BDA0002826356510000152
wherein L is1To represent the boxes of the trace results, L2A group of boxes representing i pieces of position information in the detection result, wherein i is a positive integer; IoU, the coincidence ratio between the box representing the detection result and the box representing the tracking result is calculated as follows:
Figure BDA0002826356510000153
wherein the content of the first and second substances,
Figure BDA0002826356510000154
indicating the coincidence ratio between the box representing the tracking result and the box representing the ith positioning information in the detection result, wherein j is IoU corresponding to the value of i; the formula is as follows:
Figure BDA0002826356510000155
wherein the content of the first and second substances,
Figure BDA0002826356510000156
to represent the area of the box of the trace result,
Figure BDA0002826356510000157
is the area of the box representing the ith position information in the detection result.
Further, the segmentation module is specifically configured to:
and (3) segmenting the initial coordinate position by adopting a Snake segmentation algorithm based on a depth network, wherein the final tracking result is expressed as follows:
L(x,y)＝Snake(L)；
L＝{(Max(xj),Max(yj))，(Min(xj),Min(yj))}；
wherein L is(x,y)The method comprises the following steps that a plurality of node coordinates of a segmented contour are obtained, L is a box representing a final tracking result, and the box takes the maximum value and the minimum value of a group of node coordinates in the plurality of node coordinates of the segmented contour as the upper left corner and the lower right corner; max (x)j) Is the maximum value of the abscissa, Max (y) in the coordinates of a plurality of nodes of the segmented contourj) Is the maximum value of the ordinate, Min (x), in the coordinates of a plurality of nodes of the segmented contourj) Is the minimum value of the abscissa, Min (y) in the coordinates of a plurality of nodes of the segmented contourj) The minimum value of the ordinate in the coordinates of a plurality of nodes of the segmented contour is obtained.
For the embodiments of the present invention, the description is simple because it corresponds to the above embodiments, and for the related similarities, please refer to the description in the above embodiments, and the detailed description is omitted here.
Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. The high-precision target tracking method based on the depth network and the target segmentation is characterized by comprising the following steps of:
step 1, obtaining a target image and a global image, and performing target tracking on the target area image and the global image by using a twin network algorithm to obtain a tracking result; carrying out target detection on the global image by using a target detection algorithm to obtain a detection result;
step 2, detecting the tracking drift condition of objects among the same type, and correcting the tracking result by using a correction strategy based on a target tracking track;
step 3, carrying out preferential selection on the corrected tracking result and the corrected detection result to obtain a primary coordinate position;
and 4, segmenting the initial coordinate position by utilizing a segmentation algorithm to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result.
2. The method of claim 1, wherein correcting the tracking result by using a target tracking trajectory-based correction strategy comprises:
under the condition that the number of similar objects exceeds a preset value, correcting the tracking result by utilizing a track-based scoring strategy;
under the condition that the number of the similar objects does not exceed a preset value, correcting the tracking result by utilizing a traversal correction strategy;
the traversal correction strategy comprises: when the algorithm detects that the track drifts, inquiring the positions of the objects with the highest scores of the current frame and arranging the objects in sequence, wherein the positions with the highest scores and close scores respectively correspond to the positions of similar objects of the same type; respectively traversing corresponding positions to carry out calculation, wherein the minimum calculated track drift is the corrected tracking result;
the trajectory-based scoring strategy includes: when the algorithm detects that the track drifts, a track-based correction function is added in the score calculation of the twin network algorithm to correct the score so as to inhibit the problem of tracking drift among the same classes; the method comprises the following steps of taking the track of the first five frames closest to a current frame, wherein the coordinate information is respectively as follows: l is1…L5The current frame coordinate is L6(ii) a The correction function comprises a first correction term D1And a second correction term D2Respectively as follows:
Figure FDA0002826356500000021
Figure FDA0002826356500000022
Figure FDA0002826356500000023
wherein D is1Represents the current frame L6And the previous frame L5Distance between, and average distance L of the previous five frames of the current frameAverageComparing, and correcting if the current frame has large position offset; d2Represents the current frame L6The first five frames and the last frame L5Variance comparison between the previous five frames, if presentFrame L6Correcting if the variance of the first five frames is large; omega1And ω2Weights of the first correction term and the second correction term respectively; score1And Score is the initial Score and the corrected Score calculated by the twin network, respectively; s is a correction parameter.
3. The method of claim 1, wherein selecting preferentially among the corrected tracking results and detection results comprises:
the preferred selection is carried out by utilizing a preferred selection algorithm C (L)1,L2) The following were used:
Figure FDA0002826356500000024
wherein L is1To represent the boxes of the trace results, L2A group of boxes representing i pieces of position information in the detection result, wherein i is a positive integer; IoU, the coincidence ratio between the box representing the detection result and the box representing the tracking result is calculated as follows:
Figure FDA0002826356500000025
wherein the content of the first and second substances,
Figure FDA0002826356500000026
indicating the coincidence ratio between the box representing the tracking result and the box representing the ith positioning information in the detection result, wherein j is IoU corresponding to the value of i; the formula is as follows:
Figure FDA0002826356500000027
wherein the content of the first and second substances,
Figure FDA0002826356500000028
to represent the area of the box of the trace result,
Figure FDA0002826356500000029
is the area of the box representing the ith position information in the detection result.
4. The method of claim 1, wherein the segmenting the preliminary coordinate location using a segmentation algorithm comprises: and (3) segmenting the initial coordinate position by adopting a Snake segmentation algorithm based on a depth network, wherein the final tracking result is expressed as follows:
L(x,y)＝Snake(L)；
L＝{(Max(xj),Max(yj))，(Min(xj),Min(yj))}；
wherein L is(x,y)The method comprises the following steps that a plurality of node coordinates of a segmented contour are obtained, L is a box representing a final tracking result, and the box takes the maximum value and the minimum value of a group of node coordinates in the plurality of node coordinates of the segmented contour as the upper left corner and the lower right corner; max (x)j) Is the maximum value of the abscissa, Max (y) in the coordinates of a plurality of nodes of the segmented contourj) Is the maximum value of the ordinate, Min (x), in the coordinates of a plurality of nodes of the segmented contourj) Is the minimum value of the abscissa, Min (y) in the coordinates of a plurality of nodes of the segmented contourj) The minimum value of the ordinate in the coordinates of a plurality of nodes of the segmented contour is obtained.
5. High accuracy target tracking means based on degree of depth network and target are cut apart, its characterized in that includes:
the initial positioning module is used for acquiring a target image and a global image, and performing target tracking on the target area image and the global image by utilizing a twin network algorithm to obtain a tracking result; carrying out target tracking on the global image by using a target detection algorithm to obtain a detection result;
the tracking drift suppression module is used for detecting the tracking drift condition of objects among the same type and correcting the tracking result obtained by the initial positioning module by utilizing a target tracking track-based correction strategy;
the preferred selection module is used for carrying out preferred selection in the corrected tracking result and the detection result to obtain a preliminary coordinate position;
and the segmentation module is used for segmenting the initial coordinate position obtained by the optimization selection module by utilizing a segmentation algorithm to obtain a segmentation result consisting of a plurality of contour nodes and obtain a final tracking result.
6. The apparatus of claim 5, wherein the tracking drift suppression module is specifically configured to:
under the condition that the number of similar objects exceeds a preset value, correcting the tracking result by utilizing a track-based scoring strategy;
under the condition that the number of the similar objects does not exceed a preset value, correcting the tracking result by utilizing a traversal correction strategy;
the traversal correction strategy comprises: when the algorithm detects that the track drifts, inquiring the positions of the objects with the highest scores of the current frame and arranging the objects in sequence, wherein the positions with the highest scores and close scores respectively correspond to the positions of similar objects of the same type; respectively traversing corresponding positions to carry out calculation, wherein the minimum calculated track drift is the corrected tracking result;
the trajectory-based scoring strategy includes: when the algorithm detects that the track drifts, a track-based correction function is added in the score calculation of the twin network algorithm to correct the score so as to inhibit the problem of tracking drift among the same classes; the method comprises the following steps of taking the track of the first five frames closest to a current frame, wherein the coordinate information is respectively as follows: l is1…L5The current frame coordinate is L6(ii) a The correction function comprises a first correction term D1And a second correction term D2Respectively as follows:
Figure FDA0002826356500000041
Figure FDA0002826356500000042
Figure FDA0002826356500000043
wherein D is1Represents the current frame L6And the previous frame L5Distance between, and average distance L of the previous five frames of the current frameAverageComparing, and correcting if the current frame has large position offset; d2Represents the current frame L6The first five frames and the last frame L5Variance comparison between the previous five frames if the current frame L6Correcting if the variance of the first five frames is large; omega1And ω2Weights of the first correction term and the second correction term respectively; score1And Score are the initial Score and the corrected Score calculated by the twin network, respectively, and s is the correction parameter.
7. The apparatus of claim 5, wherein the preference module is specifically configured to:
the preferred selection is carried out by utilizing a preferred selection algorithm C (L)1,L2) The following were used:
Figure FDA0002826356500000044
wherein L is1To represent the boxes of the trace results, L2A group of boxes representing i pieces of position information in the detection result, wherein i is a positive integer; IoU, the coincidence ratio between the box representing the detection result and the box representing the tracking result is calculated as follows:
Figure FDA0002826356500000051
wherein the content of the first and second substances,
Figure FDA0002826356500000052
indicating the coincidence ratio between the box representing the tracking result and the box representing the ith positioning information in the detection result, wherein j is IoU corresponding to the value of i; the formula is as follows:
Figure FDA0002826356500000053
wherein the content of the first and second substances,
Figure FDA0002826356500000054
to represent the area of the box of the trace result,
Figure FDA0002826356500000055
is the area of the box representing the ith position information in the detection result.
8. The apparatus of claim 5, wherein the segmentation module is specifically configured to:
and (3) segmenting the initial coordinate position by adopting a Snake segmentation algorithm based on a depth network, wherein the final tracking result is expressed as follows:
L(x,y)＝Snake(L)；
L＝{(Max(xj),Max(yj))，(Min(xj),Min(yj))}；
wherein L is(x,y)The method comprises the following steps that a plurality of node coordinates of a segmented contour are obtained, L is a box representing a final tracking result, and the box takes the maximum value and the minimum value of a group of node coordinates in the plurality of node coordinates of the segmented contour as the upper left corner and the lower right corner; max (x)j) Is the maximum value of the abscissa, Max (y) in the coordinates of a plurality of nodes of the segmented contourj) Is the maximum value of the ordinate, Min (x), in the coordinates of a plurality of nodes of the segmented contourj) Is the minimum value of the abscissa, Min (y) in the coordinates of a plurality of nodes of the segmented contourj) The minimum value of the ordinate in the coordinates of a plurality of nodes of the segmented contour is obtained.
CN202011449867.7A 2020-12-09 2020-12-09 Target tracking method and device based on combination of depth network and target segmentation Active CN112560651B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011449867.7A CN112560651B (en) 2020-12-09 2020-12-09 Target tracking method and device based on combination of depth network and target segmentation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011449867.7A CN112560651B (en) 2020-12-09 2020-12-09 Target tracking method and device based on combination of depth network and target segmentation

Publications (2)

Publication Number Publication Date
CN112560651A true CN112560651A (en) 2021-03-26
CN112560651B CN112560651B (en) 2023-02-03

Family

ID=75061415

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011449867.7A Active CN112560651B (en) 2020-12-09 2020-12-09 Target tracking method and device based on combination of depth network and target segmentation

Country Status (1)

Country Link
CN (1) CN112560651B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115423844A (en) * 2022-09-01 2022-12-02 北京理工大学 Target tracking method based on multi-module combination

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106127798A (en) * 2016-06-13 2016-11-16 重庆大学 Dense space-time contextual target tracking based on adaptive model
US20170103541A1 (en) * 2015-10-12 2017-04-13 Xsens Holding B.V. Integration of Inertial Tracking and Position Aiding for Motion Capture
CN106875425A (en) * 2017-01-22 2017-06-20 北京飞搜科技有限公司 A kind of multi-target tracking system and implementation method based on deep learning
CN111179307A (en) * 2019-12-16 2020-05-19 浙江工业大学 Visual target tracking method for full-volume integral and regression twin network structure
CN111497872A (en) * 2020-05-20 2020-08-07 四川万网鑫成信息科技有限公司 Method for automatically generating deviation factor to correct track drift
CN111639551A (en) * 2020-05-12 2020-09-08 华中科技大学 Online multi-target tracking method and system based on twin network and long-short term clues
CN111784747A (en) * 2020-08-13 2020-10-16 上海高重信息科技有限公司 Vehicle multi-target tracking system and method based on key point detection and correction
CN111931685A (en) * 2020-08-26 2020-11-13 北京建筑大学 Video satellite moving target detection method based on bidirectional tracking strategy
CN111986225A (en) * 2020-08-14 2020-11-24 山东大学 Multi-target tracking method and device based on angular point detection and twin network
CN112037255A (en) * 2020-08-12 2020-12-04 深圳市道通智能航空技术有限公司 Target tracking method and device

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170103541A1 (en) * 2015-10-12 2017-04-13 Xsens Holding B.V. Integration of Inertial Tracking and Position Aiding for Motion Capture
CN106127798A (en) * 2016-06-13 2016-11-16 重庆大学 Dense space-time contextual target tracking based on adaptive model
CN106875425A (en) * 2017-01-22 2017-06-20 北京飞搜科技有限公司 A kind of multi-target tracking system and implementation method based on deep learning
CN111179307A (en) * 2019-12-16 2020-05-19 浙江工业大学 Visual target tracking method for full-volume integral and regression twin network structure
CN111639551A (en) * 2020-05-12 2020-09-08 华中科技大学 Online multi-target tracking method and system based on twin network and long-short term clues
CN111497872A (en) * 2020-05-20 2020-08-07 四川万网鑫成信息科技有限公司 Method for automatically generating deviation factor to correct track drift
CN112037255A (en) * 2020-08-12 2020-12-04 深圳市道通智能航空技术有限公司 Target tracking method and device
CN111784747A (en) * 2020-08-13 2020-10-16 上海高重信息科技有限公司 Vehicle multi-target tracking system and method based on key point detection and correction
CN111986225A (en) * 2020-08-14 2020-11-24 山东大学 Multi-target tracking method and device based on angular point detection and twin network
CN111931685A (en) * 2020-08-26 2020-11-13 北京建筑大学 Video satellite moving target detection method based on bidirectional tracking strategy

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
WEN-LI ZHANG ET AL.: ""Multi-Object Tracking Algorithm for RGB-D Images Based on Asymmetric Dual Siamese Networks"", 《SENSORS》 *
余志超 等: ""结合深度轮廓特征的改进孪生网络跟踪算法"", 《西安电子科技大学学报》 *
李斌 等: ""复杂场景下深度表示的无人机目标检测算法"", 《计算机工程与应用》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115423844A (en) * 2022-09-01 2022-12-02 北京理工大学 Target tracking method based on multi-module combination

Also Published As

Publication number Publication date
CN112560651B (en) 2023-02-03

Similar Documents

Publication Publication Date Title
CN111460926B (en) Video pedestrian detection method fusing multi-target tracking clues
CN107832672B (en) Pedestrian re-identification method for designing multi-loss function by utilizing attitude information
CN109636829B (en) Multi-target tracking method based on semantic information and scene information
CN108304798B (en) Street level order event video detection method based on deep learning and motion consistency
WO2020151166A1 (en) Multi-target tracking method and device, computer device and readable storage medium
CN104615986B (en) The method that pedestrian detection is carried out to the video image of scene changes using multi-detector
Yin et al. Likelihood map fusion for visual object tracking
CN111882586B (en) Multi-actor target tracking method oriented to theater environment
CN107657625A (en) Merge the unsupervised methods of video segmentation that space-time multiple features represent
JP2006209755A (en) Method for tracing moving object inside frame sequence acquired from scene
CN111046856B (en) Parallel pose tracking and map creating method based on dynamic and static feature extraction
CN107564035B (en) Video tracking method based on important area identification and matching
CN105160649A (en) Multi-target tracking method and system based on kernel function unsupervised clustering
CN108009494A (en) A kind of intersection wireless vehicle tracking based on unmanned plane
CN112287906B (en) Template matching tracking method and system based on depth feature fusion
CN109447082A (en) A kind of scene motion Target Segmentation method, system, storage medium and equipment
Kim et al. Real-time facial feature extraction scheme using cascaded networks
CN114529584A (en) Single-target vehicle tracking method based on unmanned aerial vehicle aerial photography
CN111931571A (en) Video character target tracking method based on online enhanced detection and electronic equipment
CN112560651B (en) Target tracking method and device based on combination of depth network and target segmentation
CN111914832A (en) SLAM method of RGB-D camera in dynamic scene
CN112347967A (en) Pedestrian detection method fusing motion information in complex scene
CN111160292A (en) Human eye detection method
CN113920155A (en) Moving target tracking algorithm based on kernel correlation filtering
CN107730535A (en) A kind of cascaded infrared video tracing method of visible ray

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