Multimodal Stereo Image Registration for Pedestrian Detection

This paper presents an approach for the registration of multimodal imagery for pedestrian detection when the significant depth differences of objects in the scene preclude a global alignment assumption. Using maximization-of-mutual-information matching techniques and sliding correspondence windows over calibrated image pairs, we demonstrate successful registration of color and thermal data. We develop a robust method using disparity voting for determining the registration of each object in the scene and provide a statistically based measure for evaluating the match confidence. Testing shows successful registration in complex scenes with multiple people at different depths and levels of occlusion