Structural method for obstacle detection and terrain classification

Obstacle detection, and more generally, terrain classification are two of the most important and fundamental perception functions required for robust unmanned off-road vehicle operation. To better address these tasks, we have developed a novel method that uses multiple readings from multiple sensor modalities to compute a vector measure of the physical density of a particular world location as it appears to each sensor modality. This “density map” representation serves as a powerful discriminator for the terrain classification task. We have developed this concept into a system to characterize terrain in real time from a set of sensors on-board an autonomous vehicle by assigning each patch of terrain a type and by estimating a cost metric for the vehicle to traverse that terrain. The system is fast enough to produce these estimates in real time; on our testbed vehicle, our terrain classification system is updated at roughly 70 Hz by a variety of different ladar and radar sensors. This paper discusses our methods for modeling each sensor modality, establishing the classification system, and compensating for the fact that the sensor readings may be unsynchronized and taken from a moving vehicle. A number of experiments are presented using both a stationary platform and using the autonomous Raptor vehicle developed by SAIC for the PerceptOR program. Results indicate that this system can be used to correctly classify clear flat ground, sparse vegetation, and impenetrable vegetation, and is practical for use as a guidance system for a completely autonomous vehicle. Additionally, we have demonstrated a limited ability to use this system for more sophisticated terrain classification, such as the ability to identify metal wire fencing.