An Evaluation of Local Action Descriptors for Human Action Classification in the Presence of Occlusion

This paper examines the impact that the choice of local descriptor has on human action classifier performance in the presence of static occlusion. This question is important when applying human action classification to surveillance video that is noisy, crowded, complex and incomplete. In real-world scenarios, it is natural that a human can be occluded by an object while carrying out different actions. However, it is unclear how the performance of the proposed action descriptors are affected by the associated loss of information. In this paper, we evaluate and compare the classification performance of the state-of-art human local action descriptors in the presence of varying degrees of static occlusion. We consider four different local action descriptors: Trajectory (TRAJ), Histogram of Orientation Gradient (HOG), Histogram of Orientation Flow (HOF) and Motion Boundary Histogram (MBH). These descriptors are combined with a standard bag-of-features representation and a Support Vector Machine classifier for action recognition. We investigate the performance of these descriptors and their possible combinations with respect to varying amounts of artificial occlusion in the KTH action dataset. This preliminary investigation shows that MBH in combination with TRAJ has the best performance in the case of partial occlusion while TRAJ in combination with MBH achieves the best results in the presence of heavy occlusion.