Robust Multiple Target Tracking

ObjectVideo and University of Southern California propose to develop innovative algorithms and demonstrate a framework for robust and efficient visual tracking using Adaboost-based target detection methods, kernel-based tracker, and robust method for multi-frame data-association. The key issues of multiple target tracking are: variations in target shape and appearance due to camera viewpoint and illumination condition, non-linear target motion, self and inter-occlusion, image noise and distortion, low image contrast, and high scene clutter. We identify three main technical objectives: (i) accurate detection of targets, (ii) robust online tracking, and (iii) persistent tracking of multiple targets across multiple frames. To achieve these objectives, first, we will develop algorithms for target detection based on the AdaBoost technique, boosted-tree multi-view classifier, and using edge-based features including edgelets and histogram-of-gradients. Second, we will design a collaborative multiple-kernel algorithm for robust online tracking. Third, we will design a multi-frame algorithm for data association for persistent tracking and error correction. These algorithms address different functional requirements of a tracking system in a coordinated framework to achieve improved and efficient overall performance. We will validate the developed algorithms with thorough performance evaluation and quantitative analysis under different environments and scenarios.