Pedestrian Detection Based On Single-frame Infrared Image

Pedestrian are the main participants and casualties of the traffic, so pedestrian detection is one of the hottest topics in the domain of driver assistance systems. Besides, pedestrian detection is the main approach to get pedestrian's information of the urban road, and pedestrian detection is also the foundation of analysing pedestrian's action. Pedestrian detection also has broad applications in surveillance. Infrared camera can be used in daytime and nighttime, and it has the ability of penetrating through smoke, fog and snow. Moreover, infrared camera doesn't be disturbed by strong light, and it hasn't strong configuration of color and texture. It can achieve observing of far distance and all-weather. So pedestrian detection based on infrared image can detect pedestrian in nighttime or bad weather, and more and more scholar begin to study this problem.In this paper, we introduced the background and significance of the task, analysed the difficulties of pedestrian detection based on infrared image, and summarized the character of the infrared image and the state of pedestrian detection. The system designed in this paper concentrates on the research of the pedestrian detection in single infrared image to solve the problem of single pedestrian's detection in complex circumstance. In the paper, we used the information of infrared image's character and pedestrian's shape. In detail, the prime task of this paper concentrates on the research of the following key technologies and algorithms:1.Infrared image preprocessing:Far infrared image has low Signal-to-Noise and contrast, and the contour of the targets in it is blurry. Because of these characters, targets detection in infrared image is difficult. We need to preprocess the image before target detection. To remove the noise of the image, we use 3×3 median filtering. Median filtering is a non-linear method, and it can remove the noise not destroying the borderline of the targets. To enhance contrast of the image, we adopt joint self-adaptive threshold histogram equilibrium and power transform. Self-adaptive threshold histogram equilibrium changes the contrast of the image throught choosing threshold by itself. The contrast of the image after self-adaptive threshold histogram equilibrium is enhanced, but the whole image is bright. So we adapt power transform to reduce the image gray values. After all these processes, we remove the noise of the infrared image, and the contour of the targets become clearer.2.Regions of interest (ROIs) segmentation and ROIs elementary validating:Typical pedestrian detection system includes ROIs segmentation and pedestrian recognition, and we used this typical system in the paper. Because of the temperature of pedestrian higher than the circumstance, we use multi-level threshold segmentation to pick up hot spot in the infrared image as ROIs. Multi-level threshold segmentation chooses different thresholds at every pixel, and this method can segment the hot zone of the infrared image. After multi-level threshold segmentation, the system adapts joint morphology opening and removing small regions to remove the noise and disturbance in the binary image. Pedestrian, especially pedestrian legs have prominent vertical edge symmetry, and we validate the ROIs using this character. We also use the vertical edge symmetry to remove some regions which havn't strong vertical edge symmetry.3.Features extraction and target recognition:To recognise different targets, we should extract representative characters first.Leanness ratio and compactness are appropriate for distinguishing pedestrian and non-pedestrian, so we choose these two features to distinguish different targets in the paper. We design a BP neural network including three levels. In the paper, we choose varied pedestrian and non-pedestrian training samples to train the network using LM arithmetic. In the end, we use the classifier to recognise different targets.We can prove the classifier has high detection rate and low false rate by experiment.