Evaluation of stereo vision obstacle detection algorithms for off-road autonomous navigation

Abstract

Reliable detection of non-traversable hazards is a key requirement for off-road autonomous navigation. Under the Army Research Laboratory (ARL) Collaborative Technology Alliances (CTA) program, JPL has evaluated the performance of seven obstacle detection algorithms on a General Dynamics Robotic Systems (GDRS) surveyed obstacle course containing 21 obstacles. Stereo imagery was collected from a GDRS instrumented train traveling at 1m/s, and processed off-line with run-time passive perception software that includes: a positive obstacle detector, a negative obstacle detector, a non-traversable tree trunk detector, an excessive slope detector, a range density based obstacle detector, a multi-cue water detector, and a low-overhang detector. On the 170m course, 20 of the 21 obstacles were detected, there was complementary detection of several obstacles by multiple detectors, and there were no false obstacle detections. A detailed description of each obstacle detection algorithm and their performance on the surveyed obstacle course is presented in this paper.