Fast Pedestrian Detection Based on Saliency Detection and HOG-NMF Features

Pedestrian detection is a key ability for a variety of important applications, such as robotics, driver assistance systems and surveillance. This paper presents a fast pedestrian detection based on saliency detection and Histogram of Oriented Gradient - Non-negative Matrix Factorization (HOG-NMF) features. The regions of interest are extracted using the frequency tuned saliency detection and threshold based on entropy. A novel HOG-NMF features that reduce significantly the length of feature vector are proposed. Classification method using intersection kernel SVM offers significant improvements in accuracy over linear SVM with the same computational complexity. Experiments on INRIA dataset show that the proposed method reduces significantly runtime compared with HOG/linear SVM and HOG/RBF-SVM, achieves the satisfactory accuracy.