Real-Time Target Tracking in Long Gray-Scale Image Sequence

To solve the difficulty of real-time target tracking in long gray-scale image sequence, an improved method based on gray-scale density distribution estimation technique is proposed. A single robust M-estimation framework the matching of local gray-scale distributions is employed. To represent the target distribution model more approximately, a 3D joint spatial and gray-scale value space is defined, the reference target region can be mapped to the candidate image and vice versa, both transformations parameterized by an eight-dimensional vector. These distributions are computed with spatially oriented kernels in the 3D space and the estimation process uses only local information. The rotation, scaling and illumination change are taken into account of the distribution computation. Update these parameters iteratively, the tracking task is achieved.