A Probabilistic Approach for Human Action Recognition Using Motion Trajectories

Automatic recognition of human actions from the videos is of widespread interest for researchers as it has broad range of applications in areas such as video indexing, visual surveillance etc. Several techniques proposed for human action recognition suffer from real-time implications as they are complex and have many tunable parameters. This paper presents a simple and computationally efficient action recognition approach that extracts Harris corner points in difference of frames (DoF) and uses the magnitude of motion vectors extracted at corner points are used as weights to obtain center of mass (centroid) locations representing spatial distribution of corner points in each frame. The movement of centroid as action advances yields a motion trajectory. The sequence of quantized orientations of trajectory segments are used as temporal features for action classification that uses probabilistic state models and maximum likelihood decision making strategy. The benchmark action datasets such as KTH dataset, Weizmann dataset and UIUC complex activity dataset are used to train and test state models of each action class. This approach performed effectively by adapting to large set of intra-class variations and achieved a satisfactory classification accuracy for most confusing action classes, namely walking and running actions.