Human Action Recognition Using Trajectory-Based Spatiotemporal Descriptors

Human action recognition has gained popularity because of its wide applicability in automatic retrieval of videos of particular action using visual features. An approach is introduced for human action recognition using trajectory-based spatiotemporal descriptors. Trajectories of minimum Eigen feature points help to capture the important motion information of videos. Optical flow is used to track the feature points smoothly and to obtain robust trajectories. Descriptors are extracted around the trajectories to characterize appearance by Histogram of Oriented Gradient (HOG), motion by Motion Boundary Histogram (MBH). MBH computed from differential optical flow outperforms for videos with more camera motion. The encoding of feature vectors is performed by bag of visual features technique. SVM with nonlinear kernel is used for recognition of actions using classification. The performance of proposed approach is measured on various datasets of human action videos.