Techniques for Object Tracking: Algorithms and Benchmarks

Abstract

Visual object tracking is a fundamental computer vision task, and has a wide range of applications including video surveillance, human computer interaction, augmented reality, vehicle navigation, robotics, etc. In this dissertation, we focus on both developing robust tracking algorithms and creating benchmark datasets for evaluation and diagnosis purposes. First, to comprehensively investigate the effect of encoding color information for the visual tracking task, we develop 160 color-enhanced trackers and compile a dataset containing 128 color sequences for evaluation. We also provide detailed analysis of the results. Second, to deal with the problem that all of the current planar object tracking benchmarks are constructed in laboratory environments, we present a carefully designed planar object tracking benchmark contains 210 video sequences of 30 planar objects sampled in the wild. For each object, we shoot seven videos according to seven challenging factors. We annotate the ground truth in a semi-automatic manner to ensure the accuracy. We also evaluate two representative algorithms and provide detailed analysis of the results. Third, in order to incorporate the reliable prior knowledge that the target object in tracking must be an object other than non-object, we adapt the BING objectness measure to a specific tracking object with adaptive support vector machine. The effectiveness of the proposed adaptive objectness, named ADOBING, is generic. The performance of all the carefully selected base trackers can be improved on two popular benchmarks. Fourth, we propose a blurred target tracking algorithm using group sparse representation which can capture the natural group structure among the templates. Based on the observation that the blur templates of the same direction have similar gradient distributions, we include gradient histograms in the appearance model to further boost the performance. The resulting non-smooth optimization problem is solved with an efficient algorithm based on accelerated proximal gradient scheme. Moving vehicle detection is an important prerequisite for multiple moving vehicle tracking in wide area motion imagery. Based on the motivation that there are usually a relatively large number of vehicles in several consecutive frames along the direction of the road, we present a novel temporal context (TC) feature to capture the road context without detecting road explicitly. We evaluate TC with the CLIF dataset, and the experimental results show that TC is useful to remove false positives which are not on the road.