Computer vision in a heterogeneous software and hardware environment

Abstract

A well-developed modular, extensible vision system, based on a connectionist approach, is analyzed from a concurrent processing standpoint. This system can accurately reconstruct objects, using a set of locally derived features, from real, low-resolution-range data. The approach is highly parallel in nature. An implementation of the system in a heterogeneous multiprocessing environment is examined. Improved algorithms for low-level feature extraction are employed, including multiwindow parameter extraction and a conflict-resolution strategy. This results in improved robustness, while a simple multiprocessor environment gives a substantial speedup. Tests with real data demonstrate a factor of 10 gain in performance from mapping tasks onto appropriate hardware and software and show the potential of model-driven search in such an implementation.