ATSN: Attention-Based Temporal Segment Network for Action Recognition

In human action recognition, a reasonable video representation is still a problem to be solved. For humans, it is easy to focus on the prominent areas of the image in the video, focusing on the part of interest. Inspired by this, we proposed a deep Temporal Segment Network based on visual attention-ATSN. By lightly modifying the model structure, ATSN integrates the human attention mechanism into the Temporal Segment Networks, can effectively add a weight to the video representation features, pays attention to the beneficial regions in the features, and achieves more accurate action recognition. We conducted the Oilfield-7 dataset for human actions on the oilfield. The experimental results on HMDB51 and Oilfield-7 show that the ATSN had achieved excellent performance.