Posture-Invariant Human Detection and Tracking for Outdoor Night-Time Surveillance

Human detection and tracking from infrared image sequences has received considerable attention in many practical applications, ranging from security and surveillance to automated health-care monitoring. However, most of the systems currently reported in the literature assume that humans are in an upright standing or walking posture in the monitored scene, which may not be true in some real-world surveillance scenarios, as humans can move in other abnormal postures, such as creeping and crawling. To overcome this limitation and enable human detection even in the presence of posture changes, this paper proposes a novel system based on locating human head–shoulder Ω-like part and two legs. For tracking purposes, a particle filter and an adaptive combination of different cues, namely spatial, intensity, texture and motion velocity are used. Then, to better describe the posture of the detected human and thus enable its effective recognition over time, three different features, namely Krawtchouk moments, chain code histograms and geometry-based features are first extracted, and then fed into a dendrogram-based support vector machine classifier for posture recognition. The results of posture recognition, in combination with the tracking information, are finally exploited to analyze the behavior of the detected human in the monitored scene. The proposed system was evaluated by performing extensive experiments using several infrared image sequences taken in a real outdoor nighttime environment. The obtained results are satisfactory and demonstrate the feasibility and effectiveness of the proposed system for the automatic detection of moving humans and the analysis of their behavior.