Multi-target tracking is a key problem for many computer vision tasks, such as surveillance, animation or activity recognition. Multi-target tracking aims at reconstructing trajectories of moving targets in a dynamic scene. Owing to great advance in object detection, majority of current approaches take the tracking-by-detection strategy for multi-target tracking to get global optimum target states considering all the targets in a scene.
Still, there are a lot of challenging issues in multi-target tracking in crowded environments, such as targets hardly discernible, targets similar to the background, shadows, occlusions, illumination changes, viewpoint variations,
non-rigid targets, appearance changes, nonlinear dynamics, appearance model drifting, etc. Indeed, the performance of a tracker is greatly affected by the target density in the scene. We can call a scene sparse, when the low target density causes very few occlusions lasting a short period of time. Targets in a sparse scene move
freely toward their destination and few intra-target interactions can be observed. On the other end of the spectrum, we have a crowded scene, where no subject is fully visible and each target motion is constrained by the motion of other targets. This dissertation focuses on this crowded scene. In this scenario, there is more than a single
dominant scene motion pattern and the targets interact with one another while moving towards their destinations. We call this scene a busy scene.
To resolve correspondence ambiguities between measurement detections and target states in a busy scene, it is important to exploit all the sources of information, such as the appearance, the physical constraints. A sociological study found that 89% of people form small groups when they travel, walk, go to the cinema, work, etc. Thus small
groups are very frequent in surveillance scenarios. This thesis exploits the physical constraints based on the social context in crowded environments.
Multi-target tracking based on a global optimization method is used to get the globally optimum trajectories considering all targets in the scene. Existing global optimization based multi-target tracking assumes that each
pedestrian’s motion is independent and shows impressive results in sparse datasets. But in crowded environments,
pedestrians often occlude and interact with each other, making multi-target tracking a challenging task.
To get social context, we develop a multi-group tracking approach in a global optimization framework to get
temporally consistent group entities. In this dissertation, we consider social group behavior to mitigate ambiguities
in a multi-target tracking method using the Social Force Model (SFM) - widely used in crowd simulation applications.
To reduce ID switches among occluding objects in crowd environments, we exploit social interactions found
by the multi-group tracking method. Obtained social interactions modeled by the SFM produce the different sets
of trajectories combining repulsive, goal directed, and group attraction forces influencing the individual trajectory.
In the multi-target tracking framework, we use group dynamics by defining the prediction error for each trajectory
hypothesis obtained from group entities and extend the trajectory hypothesis space considering other individuals
within a graph structure.
With the multi-target tracking (MTT) and multi-group tracking (MGT) systems, we define a collaborative
framework in order to share MTT and MGT informations. We iteratively infer multi-target and multi-group locations.
In the established framework, we evaluate the system in synthetic and public datasets containing sparse
scenes and crowded scenes and show the effectiveness of social context in multi-target tracking in crowded environments.
다중물체 추적은 컴퓨터 비전 분야에서 가장 중요하게 다뤄지는 문제 중 하나로써 보안, 애니메이션, 움직임 인식 등 다양한 응용 분야를 갖고 있다. 다중물체 추적은 움직이는 물체의 궤적을 찾는 것을 목적으로 하며, 물체 탐지기의 발전으로 최근의 대다수의 추적기는 탐지된 결과의 대응 관계를 찾음으로써 궤적을 얻는다. 하지만 보다 복잡한 환경에서는 여전히 해결해야할 문제는 남아있다. 움직이는 물체사이의 가려짐이 있고, 외양 정보가 유사한 경우 추적을 실패하기 때문이다.
이러한 대응관계에 있어서의 불확실성을 줄이기 위해 보다 좋은 외양 모델(appearance model)을 사용하거나, 보다 정교한 움직임 모델(motion model)을 써야 한다. 본 논문에서 복잡한 환경에 초점을 두고 있고, 이러한 때에는 개개의 사람들 사이에 상호작용이 있다. 따라서 사회학적 문맥을 이용한 움직임 모델로 추적기의 정확도를 향상하도록 하였다.
물체의 궤적을 추적하기 위해 움직이는 사람사이의 반발하는 힘, 목적지를 지향하는 힘, 환경의 영향을 받는 힘, 그룹 사이의 끌어 당기는 힘을 고려하여 그룹을 탐지하였다. 그리고 이러한 사회학적 문맥을 이용하여 복잡한 환경에서도 물체의 위치를 잘 예측할 수 있도록 사회학적 문맥을 고려한 움직임 모델을 정의하고, 이 모델로 다중물체 추적기의 다중 물체의 궤적에 대한 가설을 핸들링 하였다.
이러한 방법론을 synthetic 정보와 public 데이타 상에서 정량적, 정성적으로 분석하여 관련 state-of-the-art 와 비교 검증하여 제안한 방법론의 우수성을 입증하였다.
