Research On Algorithms Of Pedestrian Detection Based On Infrared Images

Nowadays, traffic accidents are a major social problem and traffic accidents in the night account for a large proportion of these traffic accidents. For this reason, the pedestrian detection causes much attention in the research field of intelligent vehicles in recent years. When driving in the night, the light outside is not sufficient, and there are many disturbing objects in the road. In such an environment, infrared thermal imaging technology which can reflect temperature information of the object has great advantages compared with the visible light imaging technology. So this thesis focuses on the pedestrian detection algorithm based on infrared images.The pedestrians in the infrared image are regarded as research subjects and the infrared image pedestrian detection technology becomes the key point of this thesis. Samples feature extraction and classification selection these two parts are included in pedestrian detection technology. First, the infrared image should be preprocessed owing to a lot of noise points existing in the infrared images. Filtering out the noise in the image and using the infrared image detail enhancement methods to enhance the infrared image pedestrian detail informations are included in the preprocessing progress. Finally hierarchical pedestrian detection method based on Shapelet features and HOG features is conducted. The main contents of this thesis include: 1. The infrared image according to the characteristics of the infrared image is preprocessed, including denoising and infrared image enhancement this two parts. 2. A large number of infrared images are captured by using the infrared thermal imaging system, which provide good material for feature extraction of pedestrians; then the Shapelet feature is extracted on the basis of the characteristics of pedestrians, and the gradient values in four different directions are calculated for each image, then Adaboost classifier is trained to obtain a crude classifier through which this thesis can obtain candidate pedestrian area in the input image. 3. A research of hierarchical pedestrian detection algorithm based on Hog features is conducted, HOG feature of training pedestrian samples is extracted and then HOG features by RBF kernel function in SVM classifier is trained to get the final pedestrian detection classifier, detecting the pedestrian areas and locating the pedestrian in input image.The Shapelet features can well describe the gray mutations of target portion, and HOG features can describe the features of the contour and edge which can improve the consequent of target detection. The results show that the two-stage classifier improve the classification accuracy and classification speed of calssifier; and improve pedestrian detection rate to around 90% under the premise of assurancing detection speed. Finally, this thesis tests the infrared pedestrian detection algorithm in a variety of different scenarios. The results show that the proposed algorithm has better detection effect.