Automatic vehicle detection and tracking in aerial video

2016-06-21T15:03:13Z (GMT) by Xiyan Chen
This thesis is concerned with the challenging tasks of automatic and real-time vehicle detection and tracking from aerial video. The aim of this thesis is to build an automatic system that can accurately localise any vehicles that appear in aerial video frames and track the target vehicles with trackers. Vehicle detection and tracking have many applications and this has been an active area of research during recent years; however, it is still a challenge to deal with certain realistic environments. This thesis develops vehicle detection and tracking algorithms which enhance the robustness of detection and tracking beyond the existing approaches. The basis of the vehicle detection system proposed in this thesis has different object categorisation approaches, with colour and texture features in both point and area template forms. The thesis also proposes a novel Self-Learning Tracking and Detection approach, which is an extension to the existing Tracking Learning Detection (TLD) algorithm. There are a number of challenges in vehicle detection and tracking. The most difficult challenge of detection is distinguishing and clustering the target vehicle from the background objects and noises. Under certain conditions, the images captured from Unmanned Aerial Vehicles (UAVs) are also blurred; for example, turbulence may make the vehicle shake during flight. This thesis tackles these challenges by applying integrated multiple feature descriptors for real-time processing. In this thesis, three vehicle detection approaches are proposed: the HSV-GLCM feature approach, the ISM-SIFT feature approach and the FAST-HoG approach. The general vehicle detection approaches used have highly flexible implicit shape representations. They are based on training samples in both positive and negative sets and use updated classifiers to distinguish the targets. It has been found that the detection results attained by using HSV-GLCM texture features can be affected by blurring problems; the proposed detection algorithms can further segment the edges of the vehicles from the background. Using the point descriptor feature can solve the blurring problem, however, the large amount of information contained in point descriptors can lead to processing times that are too long for real-time applications. So the FAST-HoG approach combining the point feature and the shape feature is proposed. This new approach is able to speed up the process that attains the real-time performance. Finally, a detection approach using HoG with the FAST feature is also proposed. The HoG approach is widely used in object recognition, as it has a strong ability to represent the shape vector of the object. However, the original HoG feature is sensitive to the orientation of the target; this method improves the algorithm by inserting the direction vectors of the targets. For the tracking process, a novel tracking approach was proposed, an extension of the TLD algorithm, in order to track multiple targets. The extended approach upgrades the original system, which can only track a single target, which must be selected before the detection and tracking process. The greatest challenge to vehicle tracking is long-term tracking. The target object can change its appearance during the process and illumination and scale changes can also occur. The original TLD feature assumed that tracking can make errors during the tracking process, and the accumulation of these errors could cause tracking failure, so the original TLD proposed using a learning approach in between the tracking and the detection by adding a pair of inspectors (positive and negative) to constantly estimate errors. This thesis extends the TLD approach with a new detection method in order to achieve multiple-target tracking. A Forward and Backward Tracking approach has been proposed to eliminate tracking errors and other problems such as occlusion. The main purpose of the proposed tracking system is to learn the features of the targets during tracking and re-train the detection classifier for further processes. This thesis puts particular emphasis on vehicle detection and tracking in different extreme scenarios such as crowed highway vehicle detection, blurred images and changes in the appearance of the targets. Compared with currently existing detection and tracking approaches, the proposed approaches demonstrate a robust increase in accuracy in each scenario.