Deep Learning Architecture for a Wide Variety of Sensors

Intelligent sensor fusion for autonomous vehicles is key for robust navigation and obstacle detection in complex environments. Most research has focused on image-based deep learning algorithms in urban or highway scenarios. We seek to implement a system which fuses information from LiDAR, radar, and cameras using both deep learning and traditional techniques in a data-driven approach to deploy an autonomous system in an off-road environment.  This proposal lists promising approaches that have shown promise in the Phase I contract as it applies primarily to off-road environments typically encountered in military, mining, and agriculture scenarios.  These off-road environments are typically less “busy” than urban or highway environments, but pose a different set of challenges such as degraded visual environments from dust, occlusions or false positives due to foliage, or poorly-defined roads or trails.  We seek to develop and implement promising algorithms overcoming these challenges to provide robust obstacle detection and road following technology to our current and future customers. This will, in turn, increase the robustness and reliability of our deployed systems. This data-driven deep learning approach to sensor fusion in off-road environments is valuable to our technology roadmap in the autonomous vehicle market.