Trajectory feature fusion for human action recognition

Abstract

This paper addresses the problem of human action detection/recognition by investigating interest points (IP) trajectory cues and by reducing undesirable small camera motion. We first detect speed up robust feature (SURF) to segment video into frame volume (FV) that contains small actions. This segmentation relies on IP trajectory tracking. Then, for each FV, we extract optical flow of every detected SURF. Finally, a parametrization of the optical flow leads to displacement segments. These features are concatenated into a trajectory feature in order to describe the trajectory of IP upon a FV. We reduce the impact of camera motion by considering moving IPs beyond a minimum motion angle and by using motion boundary histogram (MBH). Feature-fusion based action recognition is performed to generate robust and discriminative codebook using K-mean clustering. We employ a bag-of-visual-words Support Vector Machine (SVM) approach for the learning /testing step. Through an extensive experimental evaluation carried out on the challenging UCF sports datasets, we show the efficiency of the proposed method by achieving 83.5% of accuracy.