Weighted CoHoG (W-CoHoG) feature extraction for human detection

Abstract

Human recognition techniques are used in many areas such as video surveillance, human action recognition, automobile industry for pedestrian detection, etc. The research on human recognition is widely going on and is open due to typical challenges in human detection. Histogram-based human detection methods are popular because of its better detection rate than other approaches. Histograms of oriented gradients (HOG) and co-occurrence of histogram-oriented gradients (CoHOG) are used widely for human recognition. A CoHOG is an extension of HOG and it takes a pair of orientations instead of one. Co-occurrence matrix is computed and histograms are calculated. In CoHOG, gradient directions alone are considered and magnitude is ignored. In this paper magnitude details are considered to improve detection rate. Magnitude is included to influence the feature vector to achieve better performance than the existing method. In this paper, weighted co-occurrence histograms of oriented gradients (W-CoHOG) is introduced by calculating weighted co-occurrence matrix to include magnitude factor for feature vector. Experiments are conducted on two benchmark datasets, INRIA and Chrysler pedestrian datasets. The experiment results support our approach and shows that our approach has better detection rate.