Human action recognition based on improved CoHOG-LQC

Abstract

In accordance with the miscalculation over the recognition of resemble objects in the process of human action recognition, and strong correlations between detection precision and description capability that local texture feature descriptors can achieve when acquiring the characteristics of image edge and direction, considering the defects that the low space efficiency as well as high spectral information loss of the pedestrian tracking algorithm which based on fusion among Local Binary Pattern (LBP) and Histograms of Oriented Gradient (HOG), we proposed a novel algorithm based on the fusion among Local Quantization Code (LQC) feature and Co-occurrence Histogram Oriented Gradient (CoHOG) feature for detecting passenger. Firstly, the spectral property of the image were extracted efficiently using LQC feature descriptor from image. Next, the calculation using integral image was established to withdraw edge characteristic and CoHOG features based on LQC character spectrums from the original image. For further procedure, the CoHOG edge feature are fused with them, then the fusion feature image is acquired. At last, Histogram Intersection Kernel Support Vector Machine (HIKSVM) classifiers were performed for detection and recognition. To validate the effectiveness of the algorithm, experiments are carried out on 3 public human action dataset including Weizmann, KTH and Hollywood2. The results demonstrate that the method is effective to raise accuracy and efficiency of clustering process.