Action Recognition Based On Deep Learning Framework

Action recognition is an important research topic in computer vision,meanwhile it is a very challenging hot research topic at present.Recent years,it attracts more attention due to the potential application prospect and market in automatic video analysis,video surveillance,motion event detection,human-machine interacting and virtual reality.Traditional action recognition methods contain the following steps:(1)local feature extraction and pre-processing;(2)feature pooling and quantization;(3)classifier training and action recognition.Adopting improved trajectories with Fisher vector is the main paradigm now and it obtains leading performance in several datasets.However,improved trajectories have several problems,such as the long feature extracting time,the large space to store and can't learn from data.Recently,deep learning has gained breakthrough progress in audio and image recognition and it has an overwhelmed improvement in performance over traditional methods.Nevertheless,in video action recognition field,deep learning methods have an slow progress and have't outperform traditional methods.To solve the problems that traditional methods suffered from as well as to explore the application of deep learning in action recognition,this paper propose to adopt deep learning framework for action recognition.Based on deep learning framework and new progress in action recognition,we raise two action recognition framework:(1)fuse local feature and deep feature approach;(2)convolutional feature pooling approach.The former one make adequate usage of temporal information in local feature and scene information in deep feature and fuse them to improve the action recognition performance.The latter one directly extract convolutional features from very deep network and utilize several pooling strategies,then encode them to get new feature for action recognition.Compared with traditional method,the proposed methods have the advantages of faster running,smaller disk usage and higher performance.The effectiveness is verified on several open datasets.