SIMD and mixed-mode implementations of a visual tracking algorithm

This paper describes the implementation of a feature-based visual tracking algorithm on a SIMD MasPar MP-2 and the mixed-mode PASM prototype. The sequential algorithm is introduced and a parallel version of the algorithm is described. Aspects of the mapping of this algorithm onto both machines are presented. Real image data was used to obtain timing results for all experiments. For the SIMD implementation, the performance of each subtask of the algorithm is measured as a function of submachine size used. For the subtask of finding the best match of a feature in an image, different implementations using architectural properties are compared, and a load balancing approach is examined. For the mixed-mode implementation, the best execution mode far each subtask is determined and the results are analyzed.