Inter-occlusion reasoning for human detection based on variational mean field

Detecting multiple humans in crowded scenes is challenging because the humans are often partially or even totally occluded by each other. In this paper, we propose a novel algorithm for partial inter-occlusion reasoning in human detection based on variational mean field theory. The proposed algorithm can be integrated with various part-based human detectors using different types of features, object representations, and classifiers. The algorithm takes as the input an initial set of possible human objects (hypotheses) detected using a part-based human detector. Each hypothesis is decomposed into a number of parts and the occlusion status of each part is inferred by the proposed algorithm. Specifically, initial detections (hypotheses) with spatial layout information are represented in a graphical model and the inference is formulated as an estimation of the marginal probability of the observed data in a Bayesian network. The variational mean field theory is employed as an effective estimation technique. The proposed method was evaluated on popular datasets including CAVIAR, iLIDS, and INRIA. Experimental results have shown that the proposed algorithm is not only able to detect humans under severe occlusion but also enhance the detection performance when there is no occlusion.