Research On Night-time Pedestrian Detection Based On Infrared Images

Currently, the frequent occurrences of traffic accidents at night threaten people's lives and property. The road confusion caused by the poor light at night is the main reason leading to collision and traffic accidents between vehicles and pedestrians. As the infrared image is thermal imaging, it can go through the darkness and smoke and find targets at night or in poor light circumstances. It doesn't be interfered by glare and flash, and realizes far-distance, all-weather observation. Therefore, pedestrian detection technology based on infrared images at night can realize pedestrian detection and identification at night or under adverse weather conditions, which has important practical significance in reducing the incidence of car accidents and ensuring pedestrian's safety.The pedestrian in single-frame infrared image is the study object and pedestrian detection technology at night is researched in this paper. The main research contents include the following:1. The paper firstly outlines the principles of infrared imaging technology, and then introduces the composition of night-time pedestrian detection system, as well as principles and common methods of each part of the system, analyses problem existing in current pedestrian detection technology.2. Research the common threshold segmentation algorithms used in infrared pedestrians images simultaneously. According to the distribution of infrared image's histogram, multi-cluster analysis is done to the histogram, and an improved K-means clustering center analysis algorithm is proposed, to achieve the adaptive segmentation of human body in infrared images. The method has small computational cost, is real-time and has good noise immunity and robustness.3. When the brightness of each part of the some target are not unified, which influenced by the factors of clothing's thickness and material, and part of the target has darker brightness than the background, only using threshold segmentation algorithms will cause target fragmentation. Aim to the problem, the results of threshold segmentation and canny edge detection are combined to extract the candidate target in the thesis, and a fragmental target's restoration algorithm in infrared image is proposed , by using the tensor voting algorithm to interpolate fragmental target's boundary and linking the edge according to the curve saliency map's value and direction. Experimental results show that the method can effectively restore the infrared image's fragmental targets and realize targets'complete segmentation.