MPEG CDVS Feature Trajectories for Action Recognition in Videos

Visual Action Recognition on mobile phones is a challenging problem. Mobile and wearable devices deal with power, memory, computational and hardware constraints, which mandate robust and lightweight algorithmic implementations for sophisticated vision applications, like action recognition. Compact Descriptors for Visual Search (CDVS) is an MPEG7 standard for an accelerated visual search on mobiles. In our work, we propose a mobile action recognition framework which classifies actions by tracking CDVS feature trajectories of human subjects. The proposed method capitalizes on the sparse, salient and memory efficient properties of CDVS features. Although our recognition accuracies on standard action datasets KTH, UCF50, and HMDB is not superior to the CNN based methods, our work explores and proves the feasibility of using CDVS features for action recognition.