Robust wide-area multi-camera tracking of people and vehicles to improve CCTV usage
This thesis describes work towards a more advanced multiple camera tracking system. This work was sponsored by BARCO who had developed a motion tracker (referred to as the BARCO tracker) and wanted to assess its performance, improve the tracker and explore applications especially for multi-camera systems. The overall requirement then gave rise to specific work in this project: Two trackers (the BARCO tracker and OpenCV 1.0 blobtracker) are tested using a set of datasets with a range of challenges, and their performances are quantitatively evaluated and compared. Then, the BARCO tracker has been further improved by adding three new modules: ghost elimination, shadow removal and improved Kalman filter. Afterwards, the improved tracker is used as part of a multi-camera tracking system. Also, automatic camera calibration methods are proposed to effectively calibrate a network of cameras with minimum manual support (draw lines features in the scene image) and a novel scene modelling method is proposed to overcome the limitations of previous methods. The main contributions of this work to knowledge are listed as follows: A rich set of track based metrics is proposed which allows the user to quantitatively identify specific strengths and weaknesses of an object tracking system, such as the performance of specific modules of the system or failures under specific conditions. Those metrics also allow the user to measure the improvements that have been applied to a tracking system and to compare performance of different tracking methods. For single camera tracking, new modules have been added to the BARCO tracker to improve the tracking performance and prevent specific tracking failures. A novel method is proposed by the author to identify and remove ghost objects. Another two methods are adopted from others to reduce the effect of shadow and improve the accuracy of tracking. For multiple camera tracking, a quick and efficient method is proposed for automatically calibrating multiple cameras into a single view map based on homography mapping. Then, vertical axis based approach is used to fuse detections from single camera views and Kalman filter is employed to track objects on the ground plane. Last but not least, a novel method is proposed to automatically learn a 3D non-coplanar scene model (e.g. multiple levels, stairs, and overpass) by exploiting the variation of pedestrian heights within the scene. Such method will extend the applicability of the existing multi-camera tracking algorithm to a larger variety of environments: both indoors and outdoors where objects (pedestrians and/or vehicles) are not constrained to move on a single flat ground plane.