On pedestrian detection and tracking in infrared videos

This article presents an approach for pedestrian detection and tracking from infrared imagery. The GMM background model is first deployed to separate the foreground candidates from background, then a shape describer is introduced to construct the feature vector for pedestrian candidates, and a SVM classifier is trained based on datasets generated from infrared images or manually. After detecting the pedestrian based on the SVM classifier, a multi-cues fusing algorithm is provided to facilitate the task of pedestrian tracking using both edge feature and intensity feature under the particle filter framework. Experimental results with various Infrared Video Database are reported to demonstrate the accuracy and robustness of our algorithm.

► We introduce an intensity based shape describer to construct the feature vector for pedestrian candidates.
► We fuse the intensity and edge feature based on feature distance and physical distance.
► The weight of different feature can be adaptively computed in real-time.
► Extensive tests results demonstrate the effectiveness of the proposed algorithm.