Human Action Recognition Base On Space-time Interest Points And Hashing

With the development of the Internet and the rapid growth of multimedia data,human action recognition of the video has a broad application prospective and become one of the hottest research areas in machine vision.Human action recognition based on image processing and pattern recognition is used to identify activities of human bodies in a video.In current studies,methods based on local space-time features and bag-of-words model frameworks have achieved very good results.This framework mainly consists of three components: extraction of space-time interest points and feature representation,establishment of visual bag-of-words,and training as a classifier for action recognition.In the extraction of space-time interest point and feature representation stage,this paper uses Harris3 D corner detector to detect space-time interest points,then use the Histogram of Oriented Gradients descriptor and the Histogram of Optical Flow descriptor to represent the space-time interest point features.In the bag-of-words model,this paper presents two methods for codebook learning.The first method is hashing-based method which used the Locality Sensitive Hashing and Iterative Quantization Hashing to partition the input space similar to the Euclidean-based method.The second method is based on the codebook sparse representation which uses a basis vector to represent a space-time interest point.In the final stage,this paper uses SVM classifier with the RBF kernel.Experimental results in Weizmann and Hollywood2 datasets show that the hashing-based codebook method keeps the result effective and shorten training time.In the case with a large number of local space-time features of the input,the sparse representation learning codebook method has a better performance.