Human action recognition based on local window matching

• 1、College of Computer Science and Technology, Zhejiang University of Technology Hangzhou 310023
• 2、Key Laboratory of Visual Media Intelligent Process Technology of Zhejiang Province, Hangzhou 310023

Abstract：n this paper based on the traditional Bag-of-Words, we propose a kind of human action recognition based on local window matching with synthetically considering the temporal and spatial constraints between each pose. We refer to the background substraction technique and regard pose differences characteristics of the human body as description of motion feature. The method proposed in this paper, which has been improved on the traditional quantitative feature ,object description ,the learning model based on the Bag-if-Word and other aspects. In learning model phase, we use local training to replace traditional global training, which improve the representation of feature words. In descriptor quantization processing, a new self-adaption locality linear mapping means is proposed by this paper, which replaces the conventional quantization way. In object description phase, we divide a whole action sequence into several segments and calculate expression of the features in corresponding local window matching by temporal pyramid and sliding window which are two strategies of partition segment. The composition of the whole action sequences by connecting each local feature description. Finally, in the action process of matching, we use histogram intersection to figure up the similarity of two action sequences in testing process. In the MSR Action3D database, we test the performance of the algorithm by comparing the current action recognition methods, results show that the effect is better than the previous methods of identification.

Keywords： human action recognition local window matching Bag-of-Words depth image