Human Action Recognition And Retrieve Under Big Data Environment

Human action recognition and retrieve is a basic research project in computer vision field. With the wildly used of intelligent monitoring, human-computer interaction, sports analysis and many other fields, this subject become increasingly interested in research in recent years. Traditional video human action analysis meets many challenges like occlusion, viewpoint changing, camera movement, however, large-scale data is brought to this task a new challenge. How to quickly and efficiently extract useful information in the mass of human activities in video data is a problem worthy of further study.This paper starts from human action detection in video and describes the Action Spotting detection methods, then highlights a detecting based high-level semanticfeatures extraction approach named Action Bank and used this feature to action recognition task. This method exposes two disadvantages when processing large-scale data: Firstly, when extracting feature, the detection templates are selected manually. This step not only tedious but also unsuited to be automatically deployed and used Secondly, this feature extraction algorithm is slow. It often takes tens of minutes or even a few hours to analysis a video and this speed is clearly unable to meet the needs of big data analytics.For these two issues, this thesis studied and proposed improvements.Compared to previous work of human action recognition and retrieval, the contribution of this paper is mainly reflected in the following two aspects:1. Propose a template learning Action Bank based on spectral clustering, which is a mature clustering method. Instead of tedious templates selection work in original Action Bank paper, this work uses templates obtained by machine learning. I test my approach on two relatively small dataset: KTH, UCF Sports and a large-scale human action recognition dataset: UCF50 and verify method has well performance.2. Propose product quantization Action Bank quick algorithm to solve the problem of time-consuming in the in original work. This approach uses the product quantization, a vector quantization method, converts relative distance computing between two templates to look up table, and decreases the time complexity of the algorithm. The experiment on UCF Sports dataset shows that the feature extraction time is shortened by at least one order of magnitude or more at the cost of loss a slight of recognition rate.Overall, this thesis makes two points improvement on Action Bank when dealing with large-scale video data. The improved approach can completely quickly and effectively extract features for classification and retrieval, liberate from the tedious labor of select templates, and easy to deploy and use. For large data video analysis, the two improved schemes are very meaningful.