SBIR-STTR Award

The proposal is to develop low-cost technologies for enhanced perception and terrain understanding for robotic ground vehicle navigation. We propose to combine structured lighting with stereo vision, using innovative image processing based on shape-from-shading and shadow processing. This will provide robust ability to detect and segment negative obstacles (e.g., down steps), to estimate upcoming terrain slope, to improve object detection and segmentation (including porous obstacles such as fences), and improve texture characterization. We propose to use internal self-status sensors (e.g., inertial navigation sensors, current meters, load sensors) to collect data to characterize terrain trafficability (e.g., roughness, slope, ground resistance, traction limits, slip) for path planning. The mobile robot will exectute stylized maneuvers to measure terrain trafficability characteristics. We propose to use frequency analysis feature extraction and machine learning to classify terrain based on its trafficability (supporting landmark recognition and map region localization). We propose to train machine leaning systems to predict trafficability characteristics from structured lighting/stereo vision image texture metrics and segmented-region shape features. Preliminary experiments have demonstrated the feasiblity of key elements of the proposed approach. The research products will be applicable to DoD unmanned ground vehicle programs including the Future Combat Systems (FCS) vehicles, security robots, mine clearing and unexploded ordnance removal robots. The products will have potential applicability in commercial automotive intelligent vehicle development.

Autonomous mobility requires improved terrain perception/understanding and trafficability/mobility situation awareness. The research program will develop, test and demonstrate three complementary products addressing pressing autonomous mobility needs. The first product is enhanced stereo vision with projected lighting to detect negative obstacles (e.g., gaps and down-steps), to assess lateral and longitudinal slope, and to detect wire fences. In Phase I we demonstrated negative obstacle detection using shadow isolation processing with vertically-offset cameras and light sources. The prototype will operate in the visible and NIR spectra. We will specify the design for a SWIR system. The second product is a suite of on-board non-imaging sensors to measure the vehicle-terrain interaction, with software to classify the terrain type and to estimate key mobility/trafficability characteristics (e.g., maximum speed, stopping distance, etc.). In Phase I we demonstrated terrain classification and discrimination for a small set of terrain types with a limited sensor suite. The third product is an image-based terrain classification and mobility/trafficability characterization system. It computes a feature vector for each location in the image based on a unique application of "data mining" to model texture and structure. The software is "trained" to classify terrain types and objects, and to characterize the anticipated mobility/trafficability characteristics.

Keywords:
Projected Lighting, Terrain Understanding, Vehicle Dynamics Sensors, Trinocular Stereo, Stereo Vision, Machine Learning, Shadow Isoation Process