Realistic application driven comprehensive visual target detection and tracking systems

This thesis focuses on video target detection and tracking for moving cameras and infrared cameras. It addresses the problems of tracking moving targets in moving cameras with short-term partial or full occlusion. The moving target detection and tracking in low quality outdoor infrared videos with strong clutters are explored. Some algorithms common to computer vision and video processing are proposed.;Infrared target detection and tracking by unattended outdoor cameras bring many challenges. A complete surveillance system with low false alarm is proposed to handle both small blurred moving targets and to recognize walking people who do not have significant movements on the image plane.;A fast matching method using correlation coefficients is suggested, which can find applications in target tracking or video compression. Computational complexity analysis and real world tracking examples are presented.;A robust method which can accurately detect the orientation of an image is proposed. The comparison with a widely accepted algorithm demonstrates the benefit of the proposed method. Its application in aligning binary images is tested by walking people recognition used in video surveillance systems.;Tracking moving targets in moving cameras has extensive applications. However, the camera motion needs to be compensated to isolate the motion of the moving target. A new algorithm tracking partially occluded or blurred targets is built through three modules. The affine camera motion is estimated by a fast geometric constraint global motion estimation module. A recursive least-squares filter with a forgetting factor removes disturbances. The filter provides predictions of target position and velocity so that a compact search region is formed based on the predictions. Computational efficiency and overall system performance are tested using real world video sequences.;An interval recursive least-squares (RLS) filter can overcome the parameter estimation error and input noise. This is to circumvent the potential limitation of a RLS filter. An interval RLS filter produces state estimation and prediction in narrow intervals. Simulation and performance evaluations using real world video sequences are provided to demonstrate the effectiveness of the proposed algorithm.