Occlusion prediction based on target dynamics and spatial reasoning

This paper describes a novel system, which is able to track multiple man made objects, typically vehicles moving in a natural open world scene and continue to track those vehicles even if they become occluded by a static object in the scene. The system developed interprets major structural features of an open world scene across consecutive image frames, building a symbolic map representation of the image. The system assumes no a priori knowledge of any structure within the image. It builds the map of the scene on a frame by frame basis by tracking vehicles within the image. Extracted target dynamics are used in conjunction with a high-level reasoning process to construct the scene map. The map shows regions in the image where vehicles are likely to be detected and regions where they are likely to become occluded. The tracking is complicated by the fact that the vehicles to be tracked are expected to be a large distance from the camera and as such will only occupy a small number of image pixels. Based on the scene map and current target dynamics, the system predicts target occlusion and is able to continue tracking the target even though it is not in the field of view. The system has been tested using an input sequence of vehicles moving in a complex outdoor scene, where the vehicles undergo both full and partial occlusion