The Research Of Infrared Pedestrian Detect System In Car

As part of vehicle driver assistance system, pedestrian detection is very useful in keepingurban traffic safety, therefore it become a hot research field of Intelligent Vehicle andComputer Vision. Passive infrared camera is used for pedestrian detection because it has nightvision capabilities and can be applied in both day and night. Also, the infrared image hassimple background and immunes to the information of light, texture, shadow and color.However, pedestrian is non-rigid object and can be easily interfered by heat sources in thebackground, coupled with the lower-resolution of infrared image, makes the infraredpedestrian detection an difficult and challenge task. In this paper, infrared pedestrian detectionalgorithm is studied according to the characteristics of infrared images, and a lot of work isdone to Achieve the overall vehicle infrared pedestrian detection system.Infrared pedestrian detection is mainly consist of ROI extraction and pedestrianrecognition, the main research work of this article is as fllow:1. In ROI extraction, first filter out the noise of the infrared image. Then calculatethe dual-threshold according to the distribution of infrared image backgroundpixel and segment the image with adaptive dual-threshold. Finally extract theforeground areas, and filterer out the foreground areas that do not meet thepedestrian area characteristics;2. In Pedestrian recognition, using coarse-to-fine hierarchical recognition method.First produce three probability template according to different pedestrian direction:left, right, front and rear, And roughly classify the ROIs using template matchingmethod. Then extract the HOG features of the result ROIs, and SVM classifier istrianed for further classification of ROIs. In order to improve the recognition rate,three classifier is trained for different pedestrian directions. To improve therobustness of the algorithm, region based tracking process is joined, it increases the overall recognition rate and reduce the calculation time.Finally, five infrared video are collected, and two kinds of learning Classifier isestablished for experiment purpose. Result show that our algorithm has betterperformance, the error rate of overall pedestrian recognition has dropped below1%. TheAlgorithm is eventually ported to the embedded platform DM6437and the processingspeed reachs to22frames per second, proved its worth in engineering application.