Temporal Superpixels based Human Action Localization

In this paper, we present human action localization in videos. It is a challenging task to localize actions performed in videos where the foreground areas containing video object as well as the background areas depict motion simultaneously. We set the main objective to identify action related spatio-temporal areas in a video. The proposed approach consists of two main steps. At first, we perform saturation (S plane of HSV space) based background subtraction to obtain foreground in video frames. Secondly, we compute the optical flow of video frames. Performing the superpixels segmentation of each video frame, we classify the superpixels as temporal or stationary using optical flow information and the extracted foreground. The collection of the classified temporal superpixels provide the required action localization in videos. We present qualitative as well as quantitative evaluation of our approach using UCF sports and Weizmann actions dataset. The visual results and quantitative performance measures show that the proposed approach well localizes the action areas in the taken action sequences for evaluation.