Human dectection using combined descriptor HOG-CHT
LI Yongshun;LI Yuanjiang;ZHANG Yousai;WANG Yajun;School of Electronics and Information,Jiangsu University of Science and Technology;
In order to improve the detection rate of HOG and CHT features,and further improve the detection performance,cascaded SVM classifiers with the adaptive threshold are adopted. In the feature extraction stage,HOG and sparsed CHT features are combined together to create the augmented feature vectors of the samples. Only hard samples are saved for the next stage for classification based on the output of the samples of the upper classifier. The experimental results on the INRIA Person dataset showed that HOG-CHT feature can improve the detection rate to97. 60%,with 0. 02 false positive rate. In contrast to 65. 23% recall rate for single SVM,cascaded SVMs classifier can improve the detection rate to 84. 60%,with 0 false positive rate. The proposed HOG-CHT feature has stronger power of expression and the cascaded SVM classifiers with the adaptive threshold can further improve the detection performance of the human detection system.