Multiple camera-based chamfer matching for pedestrian detection

This paper presents a vision system for detecting pedestrians using chamfer matching. We verify the effectiveness of chamfer matching for single cameras and propose a novel method for combining results from multiple views. A key insight is that making independent decisions in each camera and combining them in a higher level is prone to error. By communicating during the template matching stage, camera nodes can avoid making hard decisions. Incorporating observations from multiple cameras should in theory reduce detection error. Additionally, we provide a conceptually straightforward algorithm for building a database that maximizes the space of poses in a minimum number of templates which results in real-time performance.