Pedestrian Detection Method Based On Multiple Features Fusion Research

As the result of the increasing number of vehicles, nowadays traffic accidents occur frequently. The pedestrian safety has become an important and worldwide issue. Hence, one of the research purposes of the intelligent driver assistance system is to improve the safety of driving and reduce the traffic accidents. Meanwhile, the pedestrian detection technology plays an important role in this system.In this paper, some of the common characteristics of the human body are used, such as the Haar-Like Features, HOG features, SIFT features and the latest TED features. We make a detailed review from their concept of feature extraction method and their applications, and do experiments based on human characteristics. This paper describes the experimental classification algorithms which are used for pedestrian detection:the support vector machines and the neural network classification algorithm. Then we describe the latest research progress of these algorithms and do the experiments using them.In this paper, we propose the improved method of pedestrian detection, which fuse multi-features to solve the problem of the low degree of recognizability of the methods only using single-feature and the diversity of the human scale. The multi-features include the HOG feature and the SIFT-PCA feature. By using the Principal Component Analysis (PCA) algorithm, the algorithm of extracting SIFT-PCA feature, which transforms the feature vector space to another one, gains the most representative of the characteristic parameters and reduces the dimensions. The fused new features can get more accurate description of pedestrian information and improve the robustness of pedestrian detection.Our experimental results show that, using the AdaBoost classification algorithm, our method, by fusing multi-features, can improve the detection rate and reduce the false detection rate. Hence this method can improve the performance of human recognition and get the better results of pedestrian detection in various natural backgrounds.