Object detection and tracking from fixed and mobile platforms

Abstract

Computer Vision technology plays a fundamental role in Video Surveillance applications with the possibility to detect different categories of objects (human beings, faces, vehicles, car plates etc) in a regular stream of video recorded by surveillance cameras. Moreover, the detection process must be validated for a sufficiently long time interval (by tracking), to provide more instances of the same object/subject, and increase the rate of successful recognition/ identification (including the possibility of human supervision). The paper address the problem of object detection and tracking and the proposed solution is based on visual appearance model learning during the tracking process. Simplified HOG-like texture features are used, to achieve computationally effective solutions to be applied in practical applications of video analytics. A contrast gradient normalization solution has been adopted, with adaptive threshold estimation, to increase tracking capability along the video flow. Performance of the tracking processing chain is evaluated using the public available TLD dataset [1], to achieve quantitative and comparable data.