Abstract:
Multi-target tracking is used in many applications, such as for indoor or outdoor surveillance, intelligent vehicles, robotics, or experiments in biology. In this thesis, three types of targets are tracked, namely drosophila melanogaster, vehicles, and pedestrians. The thesis proposes specific methods for tracking the different objects. A comparative discussion of different tracking needs is provided in Chapter 6. Vision-based locomotion analysis of drosophila melanogaster is a common way for studying nervous systems. As a holometabolous insect, there are 2D crawling (larvae) and 3D freely flying (fruit fly) motions to be analysed. We conduct tracking experiments for both stages; 3D tracking in high density scenarios defines the most difficult challenge. A tracking-by-matching method is proposed for 3D tracking to alternatively operate stereo-matching and temporal-tracking processes. Experimental results prove the efficiency of the method for reducing ambiguities in both matching and tracking modules. A tracklet-based method is proposed, based on our tracking-by-matching method, which further improves the performance by suppressing fragmented trajectories. In order to test the proposed methods, a simulator is used to generate synthetic sequences with 3D ground truth. We also study tracking for vision-based driver assistance on highways. Vehicles are the most interesting targets in such scenarios. Dynamic occlusions of vehicles occur frequently, and a tracker should also be able to deal with partially-occluded vehicles. We propose an integrated occluder-occludee tracking system. It detects and tracks both fully-visible and partially-occluded vehicles, and is able to deal with dynamic occlusion patterns. Detecting and tracking a further-away occluded vehicle contributes to the understanding of overtaking processes ahead of the egovehicle. The proposed method is tested on sequences containing such overtaking scenarios. Vision-based pedestrian behaviour analysis requires pedestrian tracking. Pedestrians may change directions abruptly. We conduct 3D tracking experiments applying a unscented Kalman filter-based tracking method. The motion model suppresses the noise from detection results. By focusing on single-frame appearance information for tracking, we propose an integrated pedestrian and body-direction classifier. It classifies the input bounding boxes into the pedestrian class and a corresponding body-direction class. We also propose a method using a part-based random decision forest for body-direction classification. We generate a pedestrian-direction classification data set (available online). This data set, and two other publicly available data sets are used to test both proposed methods. The designed tracking methods vary for the considered three different targets and data recording settings (i.e. static, or moving cameras, monocular, binocular, or trinocular), but there are also general tracking principles; see Chapter 6. Keywords: 3D tracking, fruit fly, projection consistency, vehicle, occlusion handling, pedestrian, body-direction classification, Kalman filter, random decision forest, particle filter, stereo vision