Human Activity Recognition based on Summarized Semi-detailed Frame Information and Contextual Features

Abstract

In this paper, a simple but effective methodology is proposed for human activity recognition. Major focus of the work is on the design of descriptor for the video sequence. Frames are first represented by commonly used histogram of oriented gradients and histogram of flows. Using these histograms, a semi-detailed frame level descriptor is formed. It is further augmented by incorporating contextual information. Frame level descriptors are then summarized to obtain the sequence level features that are utilized in activity recognition. SVM classifier is used for recognition. Proposed methodology is simple enough and free from any tracking and intensive details. Experiment is done on the KTH dataset. Performance of the proposed descriptors and their combinations are studied. Result indicates the potential of proposed features along with the contextual information. Comparison with a number of existing systems show that the performance of the proposed methodology is comparable.