Robust detection of people in thermal imagery

Abstract

We present a new contour analysis technique to detect people in thermal imagery. Background-subtraction is first used to identify local regions-of-interest. Gradient information within each region is then combined into a contour saliency map. To extract contour fragments, a watershed-based selection algorithm is used. A path-constrained A* search is employed to complete any broken contours, from which silhouettes are formed. Results using thermal video sequences demonstrate the capability of the approach to robustly detect people across a wider range of environmental conditions than is possible with standard approaches.