KR20180077865A - Online apparatus and method for Multiple Camera Multiple Target Tracking Based on Multiple Hypothesis Tracking - Google Patents

Abstract

The present invention relates to an online apparatus and method for multi-camera multi-target tracking based on multiple hypothesis tracking (MHT), which adopts an MHT frame work based on delay determination and is for using a tracklet (partial fragment of estimated target trajectory) in two-dimensional (2D) image coordinates to reduce an operation load of the MHT for multiple cameras. The online apparatus for multi-camera multi-target tracking based on MHT comprises: a tracklet generation unit for converting detection results detected through a frame for each camera of multiple cameras into tracklets; a candidate track generation unit for generating candidate tracks which can result in tracking in a three-dimensional space through all possible temporal and spatial combinations with the tracklets converted in the tracklet generation unit; a track selection unit for selecting a candidate track with high probability of being trajectory of a target as an optimal track on the basis of probability among the candidate tracks generated in the candidate track generation unit, and continuously selecting an optimal track while feeding back the selected optimal track; and a track removal unit for checking a track by using an approximation of global track probability (AGTP) or an N scan back scheme using the length of a track, the estimated height of a detected object and the like based on the candidate tracks generated in the candidate track generation unit and the optimal track selected in the track selection unit so as to remove an unchecked track and transmit a result to the candidate track generation unit.

Description

TECHNICAL FIELD [0001] The present invention relates to an on-line apparatus and method for multi-camera tracking based on multiple hypothesis tracking,

The present invention relates to a camera tracking apparatus and method, and more particularly, to an online apparatus and method for multi-camera multi-target tracking based on multiple hypothesis tracking.

Multi-target tracking has been intensively studied as an indispensable technology in computer vision, and online and bundling methods utilizing the appearance of the target and dynamic information have recently been proposed.

However, such a method is proposed for image tracking using a single camera. When a single camera is used, visual obstacles frequently occur due to surrounding obstacles or obstructions between the objects to be tracked. This results in a deterioration of the overall tracking performance when the density of the monitored object is high or the environment is complicated. Especially in complex environments where population density is high, such as classrooms or shops, and many obstacles lead to many deadlocks.

To overcome this problem, Possegger et al. (H. Possegger, T. Mauthner, PM Roth, and H. Bischof, "Occlusion geodesics for online multi-object tracking, IEEE Conference on IEEE, 2014, pp. 1306.1313.) Inferred that the obstacles existing in the surveillance environment were covered by prior information, and proposed a method of using them in the tracking stage. In this way, robust performance is shown under fixed obstacle environments, but there is a problem in that proper performance can not be guaranteed in the absence of dynamic obstacles or prior information about obstacles.

In order to solve the problem of occlusion in such multi-target tracking, a method using a multi-camera with overlapping field of view has been proposed. In such a multi-camera multi-target tracking (MCMTT) problem, it is necessary to find a detection from the same object among the results detected by different cameras at different times. Therefore, the problem of temporal and spatial combination must be solved at the same time.

In spatial combination, we combine observed results from different cameras for the same target, and in temporal combination, the same target combines those that appear in different frames over time.

Thus, MCMTT is more complex than single camera tracking, which considers temporal combining only. For this reason, most of the recent MCMTT algorithms are based on a placement scheme to ensure good performance. However, there is a big problem that these batch-based algorithms can not be used in many practical application systems that need to process the trace in real time.

Patent Registration No. 10-1208427 (Registration date 2012.11.29) Patent Registration No. 10-1195942 (registered on October 23, 2012)

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide an on-line device and method for multi-camera multi-target tracking based on multiple hypothesis tracking, and a multiple hypothesis tracking (MHT) Work. It is also intended to use the tracklets (partial fragments of the estimated target trajectory) in two-dimensional (2D) image coordinates to reduce the computational load of the MHT on multiple cameras.

Another object of the present invention is to solve the limitations of the state-of-the-art MCMTT methods due to considering only geometric information in spatial combination. The score of the present invention is a combination of geometric information, It is aimed to have accuracy.

Another object of the present invention is to solve the high computational load of the MHT-based scheme. The probability problem solving line of the present invention combines the partitioning and conquering method using the K highest of the solutions of the previous frame with the latest heuristic algorithm .

Other objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood by those skilled in the art from the following description.

According to an aspect of the present invention, there is provided an on-line device for multi-camera multi-target tracking based on multiple hypothesis tracking according to the present invention, comprising: A candidate track generating unit for generating a candidate track that can be a tracing result in three-dimensional space through all possible temporal / spatial combinations with the tracklets transformed by the tracklet generating unit; A track selector for selecting a candidate track having a high probability of a target trajectory on the probability among the candidate tracks generated by the track generating unit as an optimum track and successively selecting an optimum track while feeding back the selected optimum track, A candidate track generated by the candidate track generating unit, and an optimum track selected by the track selecting unit, (AGTP) or an N scan back method using the estimated height of the detected object and the like, thereby removing a track whose probability is lower than the set threshold, To the candidate track generating unit.

Preferably, the tracklet is generated through a temporal combination between detections from successive frames, and the candidate track generating unit generates a candidate track through a linking process considering motion information and shape information of the generated tracklets .

Preferably, the tracklet generator includes a matching score calculator for calculating the detection result and the similarity between tracklets that have already been generated, and a matching score calculator for calculating the similarity between the tracklets based on the similarity calculated physically from the same target based on the similarity calculated by the matching score calculator And a tracklet management unit for managing the tracklets detected by the match confirmation unit. The track checking unit may be configured to detect a result of the detection and a tracklet.

Preferably, the candidate track generating unit includes a track updater for generating a new tracklet among the tracklets derived from the tracklet generator and transformed into a three-dimensional coordinate system, or updating an existing tracklet having a matching result, A tracklet combining unit that combines the tracklets generated by the same person among at least one of the converted tracklets in the tracklet generating unit 100 using at least one bundle, A track generating unit for generating a track having three-dimensional coordinates on a three-dimensional space and assigning a score to the generated track; and a track generating unit for applying the track score generated by the track updating unit and the track generating unit, And a candidate track establishing unit for establishing candidates that can be selected from the candidate tracks.

Preferably, the track selection unit comprises: an MWCP structure unit for classifying tracks having a matching group into sub-groups by using the optimum track information to be fed back; and a threshold value determination unit for determining a threshold defined as a path of an actual target for each sub- And a global hypothesis forming unit for selecting a track having a probability higher than the probability of having the selected track and feeding back the selected track to the MWCP structure unit.

Preferably, the track remover includes an Approximation of Global Track Probability (AGTP) unit for setting a global hypothesis for eliminating optimal tracks selected by the track selection unit, and an Approximation of Global Track Probability An unidentified track removing unit for removing a track having no global probability (AGTP); an N frame scanning unit for scanning an N frame based on a global hypothesis set by the AGTP unit to correct a position of an object according to the global hypothesis, And an N scan back removing unit for removing all incompatible tracks.

According to the present invention, there is provided an on-line method for multi-camera multi-target tracking based on multiple hypothesis tracking according to the present invention, comprising the steps of: (A) detecting a pedestrian in each frame, Converting the detection results into a tracklet for a person by pairing in time; and (B) performing a three-dimensional tracking through all possible temporal / spatial combinations with the transformed tracklets through a candidate track generator (C) selecting a candidate track having a probability that a probability of a trajectory of a target is higher than a preset threshold value among the candidate tracks generated through the track selector as an optimal track, Selecting optimal tracks successively while feeding back the selected optimum tracks; (D) The possibility of becoming a tracking result by using the approximate global probability (AGTP) or the N scan back method using the track length and the expected height of the detected object on the basis of the optimal track selected by the track selection unit is diagnosed Removing a track whose probability is lower than a set threshold value, and feeding back the result to the candidate track generating unit.

Preferably, the step (A) includes the steps of: calculating a result of the detection through the matching score calculator and the similarity between the tracklets already generated; and a step of calculating, based on the calculated similarity, And detecting the tracklet to be matched.

Preferably, the detection of the tracklet updates the tracklet with a valid match through the tracklet management section, and if there is no coincidence detection for the tracklet, the trackinglet is regarded as false detection and the tracklet And generating a new tracklet if there is no new tracklet.

Preferably, the step (B) includes the steps of: combining the tracklets generated by the same person among at least one of the tracklets transformed by the tracklet generator using the update result through the tracklet combining unit into at least one bundle; Generating a track having three-dimensional coordinates on a three-dimensional space on a bundle-to-bundle basis through a generation unit, applying a score to the generated track, applying an update result and a generated track score through a candidate track establishment unit And establishing candidates that can be the result.

Preferably, in the combining step, the combining is performed by any one of a spatial association, a temporal association, and a merge of the spatial combination and the temporal combination, and the spatial combination is a time- Wherein the temporal coupling is defined by the coupling between successive tracklets of the camera with the same temporal coupling.

, 두 번째는 트랙의 연속 위치의 기하학적 적합성을 고려한 연결 스코어

, 세 번째와 네 번째는 시작 스코어

와 종료 스코어

, 마지막 하나는 트랙과 관련된 탐지 간의 시각적 유사도를 나타내는 시각적 스코어

인 것을 특징으로 한다.Preferably, a score is given for each track, taking into consideration the five factors that are to be awarded to the generated track, the first being a reconstruction score < RTI ID = 0.0 >

, The second is the connection score taking into account the geometric fit of the consecutive positions of the track

, The third and fourth start score

And end score

, And the last one is a visual score indicating the visual similarity between the detections associated with the track

.

Preferably, the step (C) includes the steps of: classifying tracks having a coincident group into sub-groups using optimal track information fed back through the MWCP structure section; Selecting an optimal track having a high probability of being a path of the MWCP structure section; and feeding back the selected optimal track to the MWCP structure section.

Preferably, the step (D) includes the steps of setting a global hypothesis for eliminating optimal tracks selected through an Approximation of Global Track Probability (AGTP) unit, determining a global hypothesis based on a global hypothesis set through an unidentified track removal unit, Scanning the N frames based on the global hypothesis set by the N scan back removing unit to correct the position of the target according to the global hypothesis, And removing all tracks that are incompatible with the track.

을 검출하는 것을 특징으로 한다.Preferably, the global hypothesis is a set of tracks generated by a partition of a tracklet, which is defined as a compatible set of tracks and is generated according to compatibility between the tracks, and the track score of the global hypothesis is the highest Global hypothesis

Is detected.

The on-line device and method for multi-camera multi-target tracking based on multiple hypothesis tracking according to the present invention as described above has the effect of alleviating not only performance but also computational complexity. That is, unlike the batch mode K-best approach to a single camera case that finds tracks with high scores for each iteration and obtains K tracks over K iterations, the online system according to the present invention is designed to find K compatible track sets Which best describes multiple object tracking in each frame of multiple camera settings.

The present invention also allows our scheme to find a near optimal solution by the proposed on-line system using a partitioning and conquering algorithm based on feedback information. That is, the feedback information on the result of the past frame greatly contributes to the configuration of the tracklet, the generation of the candidate track, and the division / conquest of the MWCP configuration.

Also, since the combining condition is designed to link the current detection to the past tracklet in order to construct the tracklet, a track tree is proposed that links the current tracklet with the past candidate track to create a candidate track, and the MWCP Based on the tracking solution of the previous frame, it is effective to reduce the calculation by reconstructing it into several sub-problems.

In addition, by applying a recursive heuristic algorithm called Breakout Local Search (BSL), which is a state-of-the-art heuristic algorithm for MWCP, it is possible to reduce the amount of computation required to solve each sub-problem. In other words, BLS not only quickly finds the nearest optimal solution, but also creates multiple regional optimal solutions for a slightly modified online strategy. After resolving these sub-problems, use the resulting solution in your cut plan to remove unreliable tracks.

FIG. 1 is a block diagram illustrating a configuration of an online apparatus for multi-camera multi-target tracking based on multiple hypothesis tracking according to an embodiment of the present invention.
2 is a drawing that defines the symbols described before the description according to the embodiment of the present invention
3 is a flow chart illustrating an on-line method for multi-camera multi-target tracking based on multiple hypothesis tracking in accordance with an embodiment of the present invention.
FIGS. 4 and 5 are diagrams for explaining a process of detecting a pedestrian through the tracklet generator in FIG. 3 and converting the detection result into a tracklet for a person by pairing in time
FIGS. 6 and 7 are diagrams for explaining a process of generating a candidate track that can be a three-dimensional tracking result through the candidate track generating unit in FIG.
FIG. 8 is a view for explaining a process of removing a track based on a global hypothesis set through an unidentified track remover in FIG. 3; FIG.

Other objects, features and advantages of the present invention will become apparent from the detailed description of the embodiments with reference to the accompanying drawings.

A preferred embodiment of an on-line apparatus and method for multi-camera multi-target tracking based on multiple hypothesis tracking according to the present invention will be described with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. It is provided to let you know. Therefore, the embodiments described in the present specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention and are not intended to represent all of the technical ideas of the present invention. Therefore, various equivalents It should be understood that water and variations may be present.

Meanwhile, in the present invention, the tracklets in each view are assembled in a three-dimensional (3D) space by back projection based on the ground plane assumption. In the ground plane assumption, all targets are assumed to move on a 3D virtual plane called the ground plane. Because this home and camera network correction information can be used to obtain the 3D position for each tracklet without triangulation, the 3D coupling problem is simplified to a 2D coupling problem in the ground plane.

1 is a block diagram illustrating a configuration of an online device for multi-camera multi-target tracking based on multiple hypothesis tracking according to an embodiment of the present invention.

As shown in FIG. 1, a tracklet generator 100 converts a detection result detected through a frame into a tracklet for each camera of multiple cameras. The tracklet generator 100 converts A candidate track generating unit 200 for generating a candidate track that can be a tracing result in a three-dimensional space through all possible temporal and spatial combinations with the tracklets generated by the candidate track generating unit 200, A track selecting unit 300 for selecting a candidate track having a high probability of being a target trajectory on the basis of a probability of an optimum track and selecting an optimum track successively while feeding back the selected optimum track, (AGTP) or N (N) using the track length and the estimated height of the detected object based on the candidates generated in the track selection unit 200 and the optimum track selected by the track selector 300, A track removal unit 400 for diagnosing a possibility of being a tracking result by using a cancellation method, removing a track lower than a threshold value at which the possibility is determined, and transmitting the information to the candidate track generation unit 200 .

At this time, the tracklet is generated through a temporal combination between detections from consecutive frames, and the candidate track generating unit 200 generates a candidate track in connection with motion and shape information of the generated tracklet.

The tracklet generation unit 100 includes a matching score calculation unit 110 for calculating the detection results and the similarity between tracklets that have already been generated and the similarity calculated based on the similarity calculated by the matching score calculation unit 110 And a tracklet management unit 130 for managing the tracklets detected by the match confirmation unit 120. The tracklet management unit 130 manages the tracklets detected by the match confirmation unit 120. The tracklet management unit 130 manages the tracklets,

The candidate track generating unit 200 includes a track updater 210 for generating a new tracklet among the tracklets transformed by the tracklet generator 100 or updating an existing tracklet having a matching result, A tracklet combining unit 220 for combining the tracklets generated by the same person among at least one of the tracklets converted by the tracklet generator 100 using the result of the update unit 210, A track generating unit 230 for generating a track having three-dimensional coordinates on a three-dimensional space for each bundle combined in the tracklet combining unit 220 and giving a score to the generated track; And a candidate track establishing unit 240 for establishing candidates that can be tracked by applying the track score generated by the track generating unit 230.

The track selector 300 includes an MWCP structure unit 310 for classifying tracks having a matching group into sub-groups by using the optimum track information to be fed back, and an MWCP structure unit 310 for sub- And a global hypothesis forming unit 320 that selects an optimal track having a high probability of a target path and feeds back the selected optimum track to the MWCP structure unit 310.

The track removal unit 400 includes an Approximation of Global Track Probability (AGTP) unit 410 for setting a global hypothesis for removing the optimal tracks selected by the track selection unit 300, An unidentified track removing unit 420 for removing tracks having no approximate global probability (AGTP) among the tracks detected on the basis of the global hypothesis (AGTP is not included in the KH best global hypothesis because AGTP is defined by the KH best global hypothesis) Scanning the N frames based on the global hypothesis set by the AGTP unit 410 to correct the position of the target according to the global hypothesis and removing all tracks incompatible with the target of the fixed position, (430).

At this time, in the case of track removal in the AGTP unit 410, unidentified tracks in the same track tree are similar because the measurements are almost the same. Therefore, it is inefficient to keep all tracks in an unrecognized track tree.

That is, the AGTP unit 410 deletes all the tracks in the unacknowledged track tree except for the best track of each tree according to an Approximation of Global Track Probability (AGTP) through the unidentified track removal unit 420. However, in order to maintain the best global hypothesis, we maintain the best global hypothesis even if unspecified trajectories are not included in the best tracks. At this time, it is difficult to keep all the tracks in each tree because the number of tracks in each tree grows exponentially in the case of the confirmed track tree.

Therefore, an N scan back approach should be applied to the identified track tree to remove tracks based on the delayed decision. For this, the AGTP unit 410 finds the best global hypothesis through the N scan back removal unit 430 and then scans N frames and corrects the position of the target according to the current best global hypothesis. It then cuts out all tracks that are incompatible with the destination in the fixed location. Meanwhile, the N scan back removal unit 430 cuts the branch of each track tree from the N frames previously except for the branch including the track in the current best global hypothesis in view of the tree structure.

The operation of the on-line device for multi-camera multi-target tracking based on multiple hypothesis tracking according to the present invention will be described in detail with reference to the accompanying drawings. The same reference numerals as those in Fig. 1 designate the same members performing the same function. The symbols described before the description are defined in the following FIG. 2.

3 is a flowchart illustrating an on-line method for multi-camera multi-target tracking based on multiple hypothesis tracking according to an embodiment of the present invention.

Referring to FIG. 3, as shown in FIGS. 4 and 5, a pedestrian is detected in each frame received by each camera of the multiple cameras through the tracklet generator 100, Into a tracklet for a person in pairs by time (S10). This is performed independently for each camera.

The matching score calculation unit 110 calculates the similarity between the detection results and the tracklets that have already been generated, and then outputs the calculated similarity through the match verification unit 120 Based on whether or not the same tracklet is physically detected, thereby detecting the tracklet that is matched. And manages the detected tracklets through the tracklet management unit 130. [

A method of calculating the degree of similarity through the matching score calculation unit 110 will be described in detail.

를 정의하고, 마지막 검출이 시간 인덱스 t-1을 갖는 모든 트랙렛의 세트로서의

는 다음 수학식 1과 같다.As a newly detected detection at time t

As the set of all tracklets whose last detection has a time index t-1

Is expressed by the following equation (1).

과 검출

사이의 매칭 스코어는 다음 수학식 2 및 수학식 3과 같이 정의된다.At this time,

And detection

Is defined by the following equations (2) and (3). &Quot; (3) "

는 트랙렛

에 의해 포함된 시간 t에서의 검출이다. 그리고 상기

는 비교 간격의 길이이다. 또한 상기

및

은 후술하는 움직임 추정 기술에 기초한 양방향(순방향 및 역방향) 추적 결과이다. Here,

The tracklet

Lt; RTI ID = 0.0 > t. ≪ / RTI > And

Is the length of the comparison interval. Further,

And

(Forward and backward) tracking results based on a motion estimation technique described later.

는 시간

(+1 : 순방향, -1 : 역방향)에서

의 추정치로서, 여기서 목표의 크기가 연속 프레임 사이에서 갑자기 변하지 않는다고 가정하면,

는 다음 수학식 4와 같이 정의된다.

Time

(+1: forward, -1: reverse)

Assuming here that the size of the target does not suddenly change between consecutive frames,

Is defined by the following equation (4).

는 다음에서 설명하는 모션 추정에서

와

사이의 주요 불일치를 나타낸다. here

In the motion estimation described below

Wow

Lt; / RTI >

로부터 특징점을 추출하고 시간

에서 프레임상의 KLT 특징 추적 알고리즘으로 추적한다. In the motion estimation, the matching score calculation unit 110 calculates

And extracts feature points from time

With a KLT feature tracking algorithm on the frame.

에서 성공적으로 추적된 j번째(

) 시점과 시간

에서의 추적 결과 사이의 불일치로

를 정의하면, 불일치 세트

는 다음 수학식 5와 같이 정의된다.Time

Lt; RTI ID = 0.0 > j < / RTI >

Time and time

As a result of discrepancies between

, The inconsistency set

Is defined by the following equation (5).

는 정적 특징 점으로부터의 불일치를 거부하는 설계 매개 변수이고,

의 관계수(cardinality)가 모든 추적된 특징 지점의 수 절반보다 작은 경우

는 이동하지 않으므로

으로 결정합니다.Here,

Is a design parameter that rejects mismatches from static feature points,

Of cardinality is less than half the number of all tracked feature points

Does not move

.

Otherwise, the matching score calculation unit 110 finds a large mismatch with the largest neighbor set as shown in Equation (6).

가

에 비례하는 이웃 윈도우 크기를 나타낼 때, 다음 수학식 7과 같다.here

end

Is expressed by Equation (7). &Quot; (7) "

는

, 즉 이전 프레임에서의 트랙렛의 최종 검출로부터 얻어진다.By using forward tracking in the above,

The

, I.e. the final detection of the tracklet in the previous frame.

Next, as shown in FIG. 6 based on the similarity calculated through the match checking unit 120, a method for detecting whether a tracklet is physically identical and detecting a matching tracklet will be described in detail.

In each frame, the existing Hungarian method (HW Kuhn, " The hungarian method for the assignment problem, " Naval research logistics quarterly, vol. 97, 1955.) Matches detection and tracking.

At this time, to improve the robustness of the tracklet, the following two 3D geometric conditions are used to confirm each match.

The first condition is the distance in 3D space between the matching detection and the last detection of the matching trace. This assumes that the distance is close enough when the match is valid. To measure the 3D distance between detections, you need to know the 3D position of each detected pedestrian with a single detection. This solves the depth ambiguity arising from single detection assuming that all pedestrians are moving to a specific 3D plane.

의 이미지 좌표

를 3D 지면 평면의 좌표로 전송하는 역 투영 함수

를 정의 할 수 있다. 이때, 검출

와 트랙렛

사이의 일치가 유효하면,

와

, 마지막 검출

는 다음 수학식 8의 조건을 만족해야 한다.Using the Home and Tsai camera calibration model,

Image coordinates of

To the coordinates of the 3D ground plane

Can be defined. At this time,

And tracklet

Lt; / RTI > is valid,

Wow

, Final detection

Must satisfy the condition of the following equation (8).

는 동일한 트랙렛에서 연속 탐지 간의 허용 가능한 최대 3D 거리이며, 입력 비디오의 프레임 속도 및 대상의 평균 이동 속도와 관련된 설계 매개 변수이다.here

Is the maximum allowable 3D distance between consecutive detections on the same tracklet and is a design parameter related to the frame rate of the input video and the average moving speed of the object.

와

는 그들 사이의 일치가 유효 할 때 유사한 높이를 가질 것이다. 이에 따라, 검출된 물체의 높이를 추정하는 함수로

을 정의할 때,

와

는 다음 수학식 9의 조건을 만족 시켜서

와

사이의 일치 유효성을 보장해야 한다.The second condition is the expected height of the detected object. Assuming that the height of the target does not suddenly change,

Wow

Will have a similar height when the agreement between them is valid. Accordingly, as a function of estimating the height of the detected object

When you define,

Wow

Satisfies the following condition (9)

Wow

The validity of the agreement between the two should be guaranteed.

는 연속 프레임 사이의 목표 높이의 최대 허용 변화량에 대한 설계 매개 변수이다.here

Is a design parameter for the maximum allowable variation in target height between consecutive frames.

When the validation is completed, the tracklet is updated as a valid match item through the tracklet management unit 130, and if there is no matching detection for the tracklet, the trackinglet is terminated as a false detection. If no tracklet matches the current detection, it is created as a new tracklet.

The candidate track generating unit 200 generates a candidate track that can be a three-dimensional tracking result through all possible temporal / spatial combinations of the converted tracklets (S20).

In step S20, the track updater 210 generates a new tracklet by updating the existing tracklet with the same result or updates the existing tracklet with the same result, And combines the tracklets generated by the same person among the converted tracklets into at least one bundle by using the update result. Then, a track having three-dimensional coordinates is generated on the three-dimensional space for each combined bundle through the track generating unit 230, a score is given to the generated track, and an update result and generation Lt; RTI ID = 0.0 > track scores < / RTI >

A method of bundling tracklets through the tracklet engaging part 220 will be described in detail.

The data connection between the tracklets determines the tracklets generated on the same target. In other words, the entire tracklet is divided into several sub-sets through association. Each subset is assumed to be associated with a target or false alarm, and a false alarm is a tracklet generated by a non-target clutter.

는 전체 트랙 세트

의 모든 가능한 파티션의 집합이라고 하면, MCMTT 문제의 목표는 표적 추적을 가장 잘 설명하는 파티션

를 찾는 것이다.

The entire track set

, The goal of the MCMTT problem is to partition the partitions that best describe the target tracking

.

) 파티션은

로 정의되고, 이때, 상기

는 잘못된 경보 집합이라고 하는 잘못된 경보 집합이고, 상기

는 i번째(

) 파티션에서 k번째(

) 목표로부터 추정된 트랙렛의 집합이며 결합 집합이라고 부른다.i th (

) The partition

, ≪ / RTI >

Is a false alarm set called a false alarm set,

Is the i-th (

) Partition to the kth (

) A set of estimated tracklets from a target and is called a join set.

It should be noted that a track can be generated in a deterministic manner given a corresponding combination set. Therefore, enumerating all possible join sets is like listing all possible tracks.

(Universe) set of all possible combinations without partition index is defined as Equation (10).

를 생략한다. 그리고

의 해당 트랙을

로 표기하고

와 결합된 검출 세트를

로 표기한다. 또한 결합 집합의 생성 방식을 설명하기 위해 공간적 결합(Spatial Association), 시간적 결합(temporal association) 및 병합(merge)으로 세 가지 연산을 정의한다.Here, for convenience,

. And

The corresponding track of

And

Lt; RTI ID = 0.0 >

. In addition, three operations are defined as Spatial Association, Temporal Association, and Merge to describe how to create a binding set.

및

에 도시 된 바와 같이, 공간적 결합성은 시간적으로 겹치는 트랙렛 사이의 결합성에 의해 정의된다. 트랙렛

과

이 같은 표적인지 아닌지를 결정하기 위해, 트랙렛에서 동시 검출 사이의 거리가 섹트(Sect)에 기술된 역 투영 함수

를 갖는

에 의해 한정되는지를 검사하는 다음 수학식 11의 공간 결합성 조건을 나타낸다.The first spatial combination is the same as that of FIG. 7 (b)

And

The spatial associativity is defined by the coupling between the temporally overlapping tracklets. Tracklet

and

In order to determine whether or not such a target is present, the distance between simultaneous detections in the tracklet is determined by the inverse projection function

Having

Lt; RTI ID = 0.0 > (11) < / RTI >

및 후속하는 트랙렛

은 다음의 두 조건을 만족해야한다.And the temporal combination of the second is defined by the coupling between successive tracklets of the same camera with temporal associativity, and the same tracklet is considered to come from the same target. That is, for time combining, the preceding tracklet

And the following tracklet

Shall satisfy the following two conditions.

및

은 일시적으로 중첩되지 않아야한다. 그건, 다음 수학식 12에서 나타낸다.Condition 1: Depending on the fact that each target can generate at most one detection in each view,

And

Should not be temporarily overlapped. This is expressed by the following equation (12).

,

의 마지막 감지 및

의 첫 번째 감지인

은 3D 공간에서 충분히 근접하다. 따라서

가 한 프레임 동안 대상이 움직일 수 있는 최대 거리로 정의되는 경우, 다음 수학식 13의 조건을 갖는다.Condition 2: Because the target can not change position suddenly

,

Last detection of

The first detection of

Are close enough in 3D space. therefore

Is defined as the maximum distance the object can move during one frame, it has the following condition (13).

는 다음 수학식 14와 같이 정의된다.Finally, the third merge is a conjunctive set where the combined set of two associative relations satisfies all spatial and temporal combining conditions. These two sets are called mergeable. Based on this,

Is defined by the following equation (14).

와

의 트랙렛은 이미 모든 연결 조건을 만족하므로

의 트랙렛과

의 트랙렛 사이의 조건 만족을 검사하여 두 결합성 집합이 병합 가능한지 여부를 결정하면 된다.bracket

Wow

Tracklet already meets all the connection requirements

With a tracklet

It is only necessary to determine whether the two sets of associativity can be merged.

A method of generating a track through the track generating unit 230 will now be described in detail.

가 주어 졌을 때, 트랙 생성부(230)는 시간 t에서의

의 검출 집합

을 결정할 수 있다. 그리고

를 사용하면 시간 t에서의 트랙

의 위치

는 재구성 및 평활화의 두 단계로 추정된다.Bond set

The track generating unit 230 generates the track < RTI ID = 0.0 >

Detection set

Can be determined. And

The track at time t

Location of

Is estimated to be two stages of reconstruction and smoothing.

에서 트랙

의 추정된 3D 위치

를 생성하는 것이다. 그리고

가 공집합이 아닌 경우

는

의 기하학적 중심점으로 다음 수학식 15와 같이 정의된다.In this section, the definition of reorganization is time

Track in

Estimated 3D position of

. And

Is not an empty set

The

Is defined as the following equation (15).

에 임의의 카메라에 의해 검출되지 않으면,

이다. 이 경우에, 인접한 두 개의 재구성된 위치의 보간에 의해

를 추정한다. 즉,

와

은 각각 비어 있지 않은 감지 집합을 가진 t에서 가장 가까운 선행 시간과 다음 시간을 나타내면, 다음 시간 t에서 재구성된 3D 위치는 다음 수학식 16과 같이 정의된다.Target Time

If it is not detected by any camera,

to be. In this case, by interpolation of two adjacent reconstructed positions

. In other words,

Wow

Represents the nearest preceding time and next time in t with a nonempty detection set, the reconstructed 3D position at the next time t is defined as: < EMI ID = 16.0 >

는 다른 시간에

의 다른 3D 위치와는 독립적이다. 그러나 인접한 위치는 대상이 특정 동작에서 움직이기 때문에 상호 결합성이 높다.Found in the reconstruction phase

At another time

Lt; / RTI > 3D position. However, the adjacent positions are highly mutually connected because the object moves in a specific motion.

의 모든 개개의 재구성된 3D 위치가 발견되면, 시간 t에서

의 최종 3D 위치인

는 다음 수학식 17에 의해 얻어진다.To smoothly process the reconstructed 3D position to account for this dependency,

≪ / RTI >< RTI ID = 0.0 >

3D position of

Is obtained by the following equation (17).

는 시간 t에서 평활화된 위치를 반환하는 함수이다.here

Is a function that returns the smoothed position at time t.

On the other hand, assigning a score to a generated track suggests a score for each track, considering the five factors.

이다. 두 번째는 트랙의 연속 위치의 기하학적 적합성을 고려한 연결 스코어

이다. 세 번째와 네 번째는 시작 스코어

와 종료 스코어

이다. 이때, 각 트랙은 트랙의 시작 또는 끝 위치의 적합성을 평가합니다. 그리고 트랙이 보이는 영역 또는 입구의 경계에서 시작하거나 끝나면 트랙의 시작 또는 종료 스코어가 낮습니다. 마지막 하나는 트랙과 관련된 탐지 간의 시각적 유사도를 나타내는 시각적 스코어

이다.The first is a reconstruction score that indicates whether the track's position is the same as the detection

to be. The second is the connection score considering the geometric fit of the continuous position of the track

to be. The third and fourth are starting scores

And end score

to be. At this time, each track evaluates the suitability of the start or end position of the track. And when the track starts or ends at the boundary of the visible area or entrance, the track's start or end score is low. The last one is a visual score that represents the visual similarity between the detections associated with the track

to be.

The track score is then defined by the following equation (18).

Next, a candidate track having a high probability of being a target trajectory is selected as an optimum track on the basis of the probability among the candidate tracks generated through the track selection unit 300, and the optimal track is selected successively while feeding back the selected optimum track (S30).

In step S30, tracks having matching groups are classified into sub-groups using optimal track information fed back through the MWCP structure unit 310. [ Then, an optimal track having a high probability of being a path of an actual target for each of the classified sub-groups is selected through the global hypothesis forming unit 320, and the selected optimum track is fed back to the MWCP structure unit 310.

At this time, a method of selecting an optimal track through the global hypothesis forming unit 320 uses a BLS (Breakout Local Search) algorithm which is already known. However, this is only one embodiment, and any arbitrary solving algorithm capable of Mixed Integer Programming is available. That is, in step S3, the size of the problem based on the interpretation result in the previous frame is reduced through the feedback before using the BLS algorithm when interpreting each frame rather than having a feature in the format of the solving algorithm There are features.

Thus, by using feedback, it has the following advantages. First, in the case of the formulation, the amount of computation to solve the problem is exponentially proportional to the number of tracks considered at one time. Second, one large problem can be divided into N small problems using feedback information. have. In other words, although the original problem can be solved only into N problems, the speed is much faster because the size of each problem is very small compared to the original problem size. Third, since the problem to be solved is a tracking problem, even if the feedback information is divided into sub-groups, there is almost no performance difference from the original problem.

An approximate global probability (AGTP) or an approximate global probability (NTP) using the track length and the estimated height of the detected entity based on the candidates generated through the track remover 400 and the optimum track selected by the track selector 300, In operation S40, the possibility of the tracking result is diagnosed using the scan back method, the track having a probability lower than the threshold value is removed, and the information is transmitted to the candidate track generating unit 200.

을 찾는 것이다.In step S40, a global hypothesis for eliminating the optimal tracks selected through the Approximation of Global Track Probability (AGTP) unit 410 is set. At this time, since the global hypothesis is a set of tracks generated by a partition of a tracklet, it may be defined as a compatible track set. Thus, from a particular set of tracks, several global hypotheses can be generated according to compatibility between tracks. The goal of the tracking method is to determine the best global hypothesis among the global hypotheses

.

를 빠르게 찾는 온라인 체계를 제시한다. 이를 위해, 모든 프레임에서, t번째(

) 프레임의 전체 트랙 세트인

로부터 모든 가능한 글로벌 가설 중에서

를 찾는 최적화 문제로서 MWCP를 공식화한다.The candidate track generation unit 200 generates

The online system that quickly finds you. To do this, in every frame, the t-th (

) The entire track set of the frame

From all possible global hypotheses from

MWCP is formulated as an optimization problem.

을 사용한다. 그러나 최상의 솔루션만 현재 프레임까지 전파 할 때 로컬 최적 조건에 갇히기 쉽다. 이 문제를 해결하기 위해 최상의 솔루션뿐만 아니라 각 프레임에서

최선의 솔루션

를 찾아 다음 프레임에서 여러 MWCP를 구성하는데 사용한다.In order to reduce the calculation and improve the performance, the candidate track generating unit 200 generates an optimal solution of the previous frame

Lt; / RTI > However, only the best solution is likely to get stuck in local optimality when propagating to the current frame. To solve this problem, as well as the best solution,

The best solution

Is used to construct multiple MWCPs in the next frame.

The candidate track generating unit 200 corrects BLS, which is a state-of-the-art heuristic BLS for solving an MWCP, as follows and applies it to an on-line system for rapidly generating a multiple solution.

) 프레임에서 높은 스코어와 호환 가능한 트랙으로 구성된 글로벌 가설을 찾는

MWCPs를 구성한다. 각 MWCP는

,

로 구성되어, 이전 최상의 글로버 가설이

이라는 가정 하에 현재 최고의 글로벌 가설의 후보인 트랙 세트이다. 본 명세서에서는

을

의 관련 트랙 세트라고 부른다.First, we explain the MWCP for MCMTT,

) Find a global hypothesis that consists of high scores and compatible tracks in the frame

Configure MWCPs. Each MWCP

,

, The previous best global hypothesis

It is a track set that is currently the best global hypothesis candidate. In this specification

of

Is referred to as an associated track set.

의 트랙, (ii) (ii)의, 즉

의 자식 중 현재 프레임에서 새로 생성된 트랙, 그리고 (iii) 확인되지 않은 트랙을 포함한다.This includes three types of tracks. (i)

(Ii) (ii), that is,

A newly created track in the current frame among the children of the current frame, and (iii) an unidentified track.

프레임보다 짧은 트랙으로, 미확인 트랙은 너무 짧아 가양성 여부를 판단 할 수 없다. 따라서 미확인 트랙이 이전 솔루션에 없는 경우에도 미확인 트랙을 관련 트랙 세트에 지속적으로 삽입한다. 그러면

은 다음 수학식 19과 같이 주어진다.The unidentified track

For a track shorter than the frame, the unidentified track is too short to judge false positives. Therefore, even if the unidentified track is not in the previous solution, the unidentified track is continuously inserted into the related track set. then

Is given by the following equation (19).

에 대한 각

에 대한 MWCP는 다음 수학식 20과 같이 공식화된다.And

For each

Is formulated as: " (20) "

은

의 접합성(compatibility) 집합이다. 문제의 솔루션은

과 같은 선택 변수

로 나타낼 수 있다.At this time,

silver

And a set of compatibility of the two. The solution in question

Optional variables such as

.

을

에 대한 MWCP를 푸는 과정을 통해, 현재 프레임의 모든 MWCP에서 발견 할 수 있는 전체 솔루션을

으로 구할 수 있다.The candidate track generating unit 200 extracts each MWCP into an expanded BLS and quickly generates a plurality of local optimal solutions from a single MWCP. Therefore

of

Through the process of unpacking the MWCP for the entire frame, all the solutions that can be found in all the MWCPs

.

에서

최적 솔루션

를 선택한다. 그러면 현재 프레임의 추적 솔루션이

에서 가장 좋은 추적 솔루션이며, 다음 수학식 21로 산출된다.That is, depending on the total track score

in

Optimal solution

. Then the tracking solution for the current frame

And is calculated by the following equation (21).

일 때, 전체 트랙 집합

를 갖는 단일 MWCP를 구성하고 풀 수 있다. 그리고 MWCP를 해결하는 계산은 그것을 해결할 때 트랙 수에 기하급수적으로 비례한다.At the start frame of the input video sequence

, The entire track set

Can be configured and solved. And the calculations that solve MWCP are exponentially proportional to the number of tracks when solving them.

는

보다 훨씬 작다. 따라서 일반적으로 온라인 체계는 여러 MWCP를 해결해야만 원래 문제를 해결하는 것보다 빠르다.Usually,

The

. Therefore, online systems are generally faster than solving the original problem by solving multiple MWCPs.

Next, we describe BLS for MCMTT. BLS finds the maximum weighted clique in undirected graph with state-of-the-art heuristic algorithm. BLS is also based on iterations consisting of local search and random perturbation. Thus, if you obtain a local optimal solution from a local search, you randomly disturb the current solution to find another local optimal solution. Also, when the local search finishes with the same solution in succession, the BLS gradually increases the disturbance intensity to escape from the local branch. This is called adaptive perturbation and is the core concept of BLS.

In the online system, BLS applies the following three variants to solve each MWCP:

으로 묶어 n 번째 MWCP의 해결 결과로 반환한다.1) Multiple solutions: Original BLS currently only maintains the optimal solution. In contrast, the present invention maintains all locally optimal solutions found until the algorithm satisfies the termination condition. Then,

And returns it as the solution result of the nth MWCP.

은 그래프의 복잡성을 반영한다고 가정한다. 그래서 본 명세서에서는 최대 반복 횟수

를 다음 수학식 22와 같이 제안한다.2) Termination condition: Since there is no way to guarantee the global optimality of the solution found, the BLS uses the maximum number of iterations as a termination condition. So the maximum number of iterations is set to a huge constant to give you more opportunities to find a better solution. However, practical algorithms that require real-time requirements are difficult to handle. Therefore, it is assumed that the appropriate number of iterations is proportional to the complexity of the graph. And the size of the compatibility set

Are assumed to reflect the complexity of the graph. Therefore, in the present specification,

Is expressed by the following equation (22).

는 미리 결정된 매개 변수이며, 여기서는 모든 실험 중에 10으로 설정한다. 그러나 수학식 22는 실용적인 애플리케이션에 국한되어야 한다. 따라서 미리 정한 매개 변수인

를 사용하여 최대 반복 횟수를 적용한다.From here

Is a predetermined parameter, and is set here to 10 during all experiments. However, equation (22) should be limited to practical applications. Therefore,

The maximum number of iterations is applied.

을

으로 설정하고 로컬 검색을 수행하여 초기 솔루션을 개선한다. 그러나 추적이 진행됨에 따라 트랙 간의 호환성이 변경 될 수 있으므로 실행 불가능할 때 로컬 검색 전에

을 복구해야 한다.3) Initial solution: BLS creates an initial solution by random selection of compatible tracks and is natural when there is no prior information. However, since the object moves smoothly between consecutive frames in the online MCMTT, the solution of the previous frame can be a powerful advance information about the current MWCP. therefore

of

And perform a local search to improve the initial solution. However, as the track progresses, compatibility between tracks may change, so if it is not feasible,

.

의 복구는 다음과 같은 방식으로 수행된다.

Recovery is performed in the following manner.

를

으로 설정한다. 그런 다음

에서 가장 높은 트랙 스코어를 가진 트랙을

에 삽입하고

를

로 업데이트 한다. 그리고

가 빈 세트가 될 때까지 이러한 삽입 및 업데이트를 반복한다.First, the candidate track of the initial solution

To

. after that

Tracks with the highest track score in

And

To

. And

≪ / RTI > repeats these inserts and updates until it becomes an empty set.

As shown in FIG. 8 based on the global hypothesis set through the unconfirmed track removal unit 420, it is assumed that there is no approximate global probability (AGTP) among the detected tracks (AGTP is defined by the KH best global hypothesis Because it does not belong to the KH best global hypothesis). Also, N frames are scanned on the basis of the global hypothesis set through the N scan back removal unit 430 to correct the position of the object according to the global hypothesis, and all tracks incompatible with the object of the fixed position are removed. At this time, the approximate global probability (AGTP) and N scan back cancellation are already known contents, and detailed description thereof will be omitted.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. It will be apparent to those skilled in the art that various modifications may be made without departing from the scope of the present invention. Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

Claims (15)

A tracklet generator for converting a detection result detected through a frame into a tracklet for each camera of the multiple cameras,
A candidate track generating unit for generating a candidate track that can be traced in three-dimensional space through all possible temporal / spatial combinations with the tracklets transformed by the tracklet generating unit;
A track selecting unit for selecting a candidate track having a high probability of being a target trajectory on the probability among the candidate tracks generated by the candidate track generating unit as an optimum track and continuously selecting an optimum track while feeding back the selected optimum track, Wow,
(AGTP) or an N scan back method using the track length and the expected height of the detected object based on the candidates generated by the candidate track generating unit and the optimal track selected by the track selecting unit And a track eliminating unit for diagnosing a possibility of being a tracking result, removing a track lower than a threshold value at which the possibility is set, and transmitting the result to the candidate track generating unit. On-line device for multi-target tracking. The method according to claim 1,
Wherein the tracklet is generated through a temporal combination between detections from successive frames and the candidate track generator generates a candidate track in connection with motion and shape information of the generated tracklet. Multi - camera based on - line device for multi - target tracking. 2. The apparatus of claim 1, wherein the tracklet generator
A matching score calculation unit for calculating a result of each detection and the similarity between already existing tracklets,
A match check unit for checking whether a tracklet is physically the same based on the similarity calculated by the matching score calculator and detecting a matching tracklet,
And a tracklet management unit for managing the tracklets detected by the match verification unit. The on-line apparatus for multiple camera multi-target tracking based on multiple hypothesis tracking. The apparatus of claim 1, wherein the candidate track generating unit
A track updater for generating a new tracklet among the transformed tracklets in the tracklet generating unit or updating an existing tracklet having a matching result,
A tracklet combining unit for combining the tracklets generated by the same person among at least one of the tracklets converted by the tracklet generating unit using the result of the track updating unit,
A track generating unit for generating a track having three-dimensional coordinates on a three-dimensional space for each bundle combined at the tracklet combining unit and giving a score to the generated track;
And a candidate track establishment unit for establishing candidates that can be tracked by applying the track score generated by the track update unit and the track update unit. The multi-camera multi-target tracking based multi- On-line device for. The apparatus of claim 1, wherein the track selector
An MWCP structure unit for classifying tracks having coincident groups into sub-groups using optimum track information to be fed back,
And a global hypothesis forming unit for selecting a track higher than a probability having a threshold defined as a path of an actual target per sub-group classified by the MWCP structure unit and feeding back the selected track to the MWCP structure unit Multi - camera based on multiple hypothesis tracking On - line device for multi - target tracking. The apparatus of claim 1, wherein the track removing unit
An Approximation of Global Track Probability (AGTP) unit for setting a global hypothesis for eliminating the optimum tracks selected by the track selection unit,
An unidentified track removal unit for removing a track having no approximate global probability (AGTP) among tracks detected based on a global hypothesis set by the AGTP unit;
And an N scan back removing unit for scanning N frames based on the global hypothesis set by the AGTP unit and correcting the position of the object according to the global hypothesis and removing all tracks incompatible with the object at the fixed position And an on-line device for multi-camera multi-target tracking based on multiple hypothesis tracking. (A) detecting a pedestrian in each frame of an image received from each camera of the multiple cameras, and converting the detected result into a tracklet for a person in pairs by time;
(B) generating a candidate track that can be a three-dimensional tracking result through all possible temporal / spatial combinations with the transformed tracklets through a candidate track generation unit;
(C) a candidate track having a probability that a trajectory of a target is higher than a set threshold value is selected as an optimum track on the basis of the probability among the candidate tracks generated through the track selection unit, and the optimal track is selected successively while feeding back the selected optimum track , ≪ / RTI &
(D) an approximate global probability (AGTP) or an N scan back method using the track length and the expected height of the detected object based on the candidates generated through the track eliminating unit and the optimum track selected by the track selecting unit And a step of diagnosing a possibility of becoming a tracking result, removing a track lower than a threshold value at which the possibility is set, and feeding back the result to the candidate track generating unit. An on-line method for target tracking. 8. The method of claim 7, wherein step (A)
Calculating a result of each detection and similarity between already existing tracklets through the matching score calculator,
And checking whether the same tracklet is physically the same based on the calculated similarity through a match check unit and detecting a tracklet matched with the tracklet. The on-line method for multiple camera multi-target tracking based on multiple hypothesis tracking. 9. The method of claim 8,
The detection of the tracklet updates the tracklet as a valid match through the tracklet management section, and if there is no coincidence detection for the tracklet, the trackinglet is regarded as a false detection and there is no tracklet matching the current detection And generating the new tracklet as a new tracklet. ≪ RTI ID = 0.0 > [0002] < / RTI > 8. The method of claim 7, wherein step (B)
Combining the tracklets generated by the same person among at least one of the transformed tracklets in the tracklet generating unit using the update result through the tracklet combining unit;
Generating a track having three-dimensional coordinates on a three-dimensional space for each bundle through the track generating unit;
Generating a score for the generated track, and applying the update result and the generated track score to the candidate track establishment unit through the candidate track establishment unit to establish candidates that can be tracked results. An on-line method for camera multi-target tracking. 11. The method of claim 10,
In the combining step, the combining is performed by any one of a spatial association, a temporal association, and a merge of the spatial combination and the temporal combination,
Wherein the temporal combination is defined by the coupling between consecutive tracklets of the same camera with temporal connectivity, wherein the temporal combination is defined by the temporal coupling between the temporal overlapping tracklets. ≪ RTI ID = 0.0 > An on-line method for multi-target tracking. 11. The method of claim 10,
Assigning a score to the generated track gives a score for each track, taking into account the five factors,
The first is a reconstruction score that indicates whether the track's position is the same as the detection

, The second is the connection score taking into account the geometric fit of the consecutive positions of the track

, The third and fourth start score

And end score

, And the last one is a visual score indicating the visual similarity between the detections associated with the track

Wherein the multiple tracking target tracking is based on multiple hypothesis tracking. 8. The method of claim 7, wherein step (C)
Grouping the tracks having a matching group into a sub-group using optimal track information fed back through the MWCP structure section;
Selecting an optimal track having a high probability of a path of an actual target for each of the classified sub-groups through a global hypothesis forming unit;
And feedbacking the selected optimal track to the MWCP structure. ≪ Desc / Clms Page number 21 > 8. The method of claim 7, wherein step (D)
Setting a global hypothesis for eliminating optimal tracks selected through an Approximation of Global Track Probability (AGTP)
Removing tracks having no approximate global probability (AGTP) among the detected tracks based on a global hypothesis set through an unidentified track remover;
Scanning the N frames based on the global hypothesis set through the N scan back removing unit to correct the position of the object according to the global hypothesis and removing all tracks incompatible with the object at the fixed position Multi - camera based on multiple hypothesis tracking with on - line tracking for multi - target tracking. 15. The method of claim 14,
The global hypothesis is a set of tracks generated by a partition of a tracklet, defined as a set of compatible tracks, generated according to compatibility between tracks,
Based on the track score defined in the previous section, the global hypothesis with the highest track score

Based tracking for multi-camera multi-target tracking based on multiple hypothesis tracking.

Images