Reading Signs in Context to Support Perception-Based Autonomous Navigation

"DoD has placed high priority on the development of autonomous systems capable of operating in complex human infrastructures (Roadways, depots, buildings, urban areas, etc.). These Infrastructures have evolved to take advantage of the human perceptualcapability by providing navigational, safety, and informational data through the use of signs. To be successful, robots developed to operate in these environments must recognize and interpret various textual and graphical content from signs, labels, andplaques and reason about the extracted information to update navigation estimates, modify control behaviors, react to unsafe conditions, and trigger discrete events. Our approach builds upon previous R&D to provide a real-time architecture for detecting,tracking, and recognizing signs from dynamic video imagery. Our approach detects signs by combining visual cues consisting of color, shape, and text. Signs are tracked to improve detection accuracy and provide resiliency to occlusions and environmentaleffects. The orientation of the sign is determined and the imagery rectified for recognition processing. Text and graphics characters are extracted using a combination of optical character recognition and syntactic/semantic parsing. Feedback loops areprovided to resolve ambiguities, improve recognition rates, and reduce false detections. Our Phase I program will result in thorough testing of the algorithms and an end-to-end feasibility demonstr