Open Collections

Abstract

For real-time motion tracking the computational requirements suggest a parallel processing solution be followed. This thesis describes the implementation of a parallel motion tracking system. Focus is placed on the resulting system as well as some of the factors involved when performing a conversion from an existing sequential system. Different parallelization methods are compared along with many of the other related issues involved in implementing vision systems on multiple processors. The model-based motion tracking system uses detected line segments from the image as features, and tracks these features with a least-squares solution to match the model with the correct feature configuration. A pipelined multiprocessor system was constructed which is able to achieve a high level of throughput using many features of the TMS320C40 parallel processors. The system consists of six C40 processors which communicate through ports controlled by a software router. A "handshaking" scheme is employed to coordinate data transfers between processors. The motion tracking system was able to track moving objects at the rate of about 4 frames per second.