A Fast Action Recognition Method with Cascaded Networks

Action recognition emerges in an endless stream and has become a hot topic in computer vision research. Obviously, fast response on recognizing video is highly desirable for the consideration of practical large-scale applications, like smart home. In our view, input frames and network structure play important roles on improving the efficiency of the action recognition. For that, a novel three-stage action recognition cascaded network architecture is proposed in this paper that consists of SelectNet, CaptureNet, and 2D+2.5D low-rank Net. The SelectNet which is a relatively-shallow network designed to filter informative video frames can guarantee all the selected frames contain meaningful information. For CaptureNet, we adopt object detection algorithm to capture object area which can avoid the influence of the clutter background. 2D+2.5D low-rank Net is an efficient network for performing action recognition which can extremely reduce computational cost. Extensive experiments on benchmark dataset UCF101 have been conducted and clearly shown that our approach yields the similar accuracy to most state-of-the-art methods at much faster processing speed.