Leveraging textural features for recognizing actions in low quality videos

Abstract

Human action recognition is a well researched problem, which is considerably more challenging when video quality is poor. In this paper, we investigate human action recognition in low quality videos by leveraging the robustness of textural features to better characterize actions, instead of relying on shape and motion features may fail under noisy conditions. To accommodate videos, texture descriptors are extended to three orthogonal planes (TOP) to extract spatio-temporal features. Extensive experiments were conducted on lowquality versions of theKTH and HMDB51 datasets to evaluate the performance of our proposed approaches against standard baselines. Experimental results and further analysis demonstrated the usefulness of textural features in improving the capability of recognizing human actions from low quality videos.