Reasearch Of Pedestrian Detection Technology Based On Multi-Features

Pedestrian detection is a hot research topic in the fields of image processing and computer vision. It has a lot of important applications in intelligent transportation, human action analysis, robot development, video surveillance and so on. Pedestrian has the characteristics of both rigid and non-rigid objects, and is affected easily by clothing, pose, lighting conditions, background and patial occlusion. So its detection is also a challenging research problem. Pedestrian detection has both academic and practical value.A pedestrian detection system consists of such components as image capturing, feature extraction, classifier training and testing. This thesis makes a comprehensive survey of the state-of-the-art of pedestrian detection, analyzes the characterisitics of some commonly used pedestrian databases and the difficulties of pedestrian detection research, and focuses on the study of feature representation of pedestrian. The main work and innovation of this thesis are as follows:(1) Inder to overcome the drawbacks with the commonly used Histogram of Oriented Gradients (HOG) feature, i.e. high dimention, high complexity to compute, high sensitivity to noise, and low ability for texture feature description, we propose the Sqrt-Local Ternary Patterns (Sqrt-LTP) for pedestrian detection. This idea is motivated by the good merits of the Local Binary Patterns (LBP). We firstly introduce the Local Ternary Patterns (LTP) into pedestrian detection, and then by calculating the square root of each code of LTP, we achieve the improved LTP, called Sqrt-LTP. Experimental results on INRIA pedestrian detection database show that the LTP based pedestrian detection method achieves higher performance than HOG based method, and the Sqrt-LTP feature can further improved the pedestrian detection results.(2) In order to solve the problem that single feature connot fully depict the image information, a novel pedestrian detection method based on multiple features is presented. Our method is implemented by fusing the commonly used HOG feature and our proposed Sqrt-LTP feature. The former can well depicts the edge information of imagse, and the latter can well represent the texture information of images. Experimental results on INRIA pedestrian detection database show that the performance of the proposed HOG+Sqrt-LTP feature based method ourperforms the single HOG feature or Sqrt-LTP feature based method.