Institute for Brain and Neural Systems

摘要

Detection and identification of partially occluded targets in complex scenes becomes an increasingly important task in light of the latest developments in urban warfare. The construction of a system that can automatically identify selected targets or direct soldiers attention to the locations that may contain suspicious activity can be of great use not only as a tool that can reduce the cognitive workload of the soldier but also as a tool that can alert the soldier to possible threats. Identifying a target in a complex scene is a challenging problem that incorporates several important aspects of vision including: translation and scale invariant recognition, robustness to noise and ability to cope with significant variations in lighting conditions. Identifying an occluded target adds another layer of complexity and this problem can be extremely difficult even for humans. Motion information can be of great help in providing an initial figure-ground segmentation. However, in many situations motion information is not available. In addition, if the input to the system is a video stream then the requirement that the system works in real-time often precludes the use of more sophisticated but computationally involved techniques.