A joint estimation of head and body orientation cues in surveillance video

The automatic analysis and understanding of behavior and interactions is a crucial task in the design of socially intelligent video surveillance systems. Such an analysis often relies on the extraction of people behavioral cues, amongst which body pose and head pose are probably the most important ones. In this paper, we propose an approach that jointly estimates these two cues from surveillance video. Given a human track, our algorithm works in two steps. First, a per-frame analysis is conducted, in which the head is localized, head and body features are extracted, and their likelihoods under different poses is evaluated. These likelihoods are then fused within a temporal filtering framework that jointly estimate the body position, body pose and head pose by taking advantage of the soft couplings between body position (movement direction), body pose and head pose. Quantitative as well as qualitative experiments show the benefit of several aspects of our approach and in particular the benefit of the joint estimation framework for tracking the behavior cues. Further analysis of behavior and interaction could then be conducted based on the output of our system.