Abstract
This paper presents and examines a hardware implementation of a high speed Iterative Closest Point (ICP) based object tracking system, which uses stereo vision disparities as input. Custom field programmable gate array (FPGA) hardware has been designed to handle the inherent bottlenecks that result from the large input and processing bandwidths of the range data. The custom hardware has been implemented and tested on various objects, using both software simulation and hardware tests. Results indicate that the tracker is able to successfully track freeform objects along arbitrary paths at rates of over 200 frames-per-second. Tracking errors are low, in spite of substantial sensor and stereo extraction noise. The tracker is able to track linear paths within 1.57 mm and 2.80 degrees and gracefully degrades under occlusion. This high speed hardware implementation with 16 parallel nearest neighbor units has a five times speed improvement when compared to a software k-d tree implementation.