Adaptive object segmentation and tracking

Efficient tracking of deformable objects moving with variable velocities is an important current research problem. In this thesis a robust tracking model is proposed for the automatic detection, recognition and tracking of target objects which are subject to variable orientations and velocities and are viewed under variable ambient lighting conditions. The tracking model can be applied to efficiently track fast moving vehicles and other objects in various complex scenarios. The tracking model is evaluated on both colour visible band and infra-red band video sequences acquired from the air by the Sussex police helicopter and other collaborators. The observations made validate the improved performance of the model over existing methods. The thesis is divided in three major sections. The first section details the development of an enhanced active contour for object segmentation. The second section describes an implementation of a global active contour orientation model. The third section describes the tracking model and assesses it performance on the aerial video sequences. In the first part of the thesis an enhanced active contour snake model using the difference of Gaussian (DoG) filter is reported and discussed in detail. An acquisition method based on the enhanced active contour method developed that can assist the proposed tracking system is tested. The active contour model is further enhanced by the use of a disambiguation framework designed to assist multiple object segmentation which is used to demonstrate that the enhanced active contour model can be used for robust multiple object segmentation and tracking. The active contour model developed not only facilitates the efficient update of the tracking filter but also decreases the latency involved in tracking targets in real-time. As far as computational effort is concerned, the active contour model presented improves the computational cost by 85% compared to existing active contour models. The second part of the thesis introduces the global active contour orientation (GACO) technique for statistical measurement of contoured object orientation. It is an overall object orientation measurement method which uses the proposed active contour model along with statistical measurement techniques. The use of the GACO technique, incorporating the active contour model, to measure object orientation angle is discussed in detail. A real-time door surveillance application based on the GACO technique is developed and evaluated on the i-LIDS door surveillance dataset provided by the UK Home Office. The performance results demonstrate the use of GACO to evaluate the door surveillance dataset gives a success rate of 92%. Finally, a combined approach involving the proposed active contour model and an optimal trade-off maximum average correlation height (OT-MACH) filter for tracking is presented. The implementation of methods for controlling the area of support of the OT-MACH filter is discussed in detail. The proposed active contour method as the area of support for the OT-MACH filter is shown to significantly improve the performance of the OT-MACH filter's ability to track vehicles moving within highly cluttered visible and infra-red band video sequences