A study of relative motion point trajectories for action recognition

Trajectories extracted by previous methods for human action recognition contain irrelevant changes, and the Orientation-Magnitude descriptors of their shapes lack the robustness to camera motion. To solve these problems, action recognition by tracking salient relative motion points is proposed in this paper. Firstly, motion boundary detector which suppresses the camera constant motion is utilized to extract motion features. After processing the detected boundaries by the adaptive threshold, the super-pixels that contain salient points are defined as relative motion regions. Then tracking the points within super-pixels is to generate trajectories. For the trajectory shape, the pre-defined orientation assignments with coarse-to-fine quantization levels are used to produce orientation statistics. Finally, the descriptors of oriented gradient, motion boundary, oriented statistic and their combination are adopted to represent action videos, respectively. On the benchmark KTH and UCF-sports action datasets, experimental results show that the extracted trajectories can describe the movement process of object. Compared with the conventional algorithms, our method with multiple kernel learning obtains good performance.