Research On Pedestrian Detection Based On The On-board Infrared Images

With the rapid development of the automobile industry,the problems of traffic security are severe and especially the nighttime traffic security has become one of the hot topics of social concern.The on-board infrared camera can effectively capture the nighttime road picture,thereby extending the observation range of drivers and reducing the occurrence of traffic accidents.Therefore,the research of pedestrian detection based on infrared images is of great theoretical significance and application value for improving traffic security.The main research work of this paper includes the following respects:1.Infrared image acquisition and preprocessing were studied.On the basis of analyzing the characteristics of infrared images,infrared images are collected and filtered,and the best filtering effect is obtained through median filtering,which ensures the quality of infrared images.2.The pedestrian detection algorithm based on feature extraction and machine learning was studied.The histogram of oriented gradient(HOG)and the support vector machine(SVM)were combined to carry out the pedestrian detection experiment.Experiments show that the algorithm has the accuracy of 76.5%but the poor real-time performance.3.The pedestrian detection algorithm based on deep convolutional neural network was studied.For the classical method cannot meet the requirement of real-time detection and accuracy on far-infrared pedestrian detection,a far-infrared pedestrian detection method based on improved YOLO,which is a deep learning based model,is proposed.The high-quality night far-infrared images collected in this paper are used as a sample to perform training one more time on network and adjusting the network training strategies,optimizing the filtering rules for candidate boxes,changing the input resolution of the network to improve the training effect.Besides,an adaptive resolution model based on the speed of the vehicle was proposed to give full play to the maximum performance of the detection system.Experiments show that the accuracy of the proposed method is better than the traditional methods from 76.5%to 91.2%and the method meets the requirement of vehicle real-time detection from 0.1493 fps to 25.4 fps.4.Image stitching technology based on binocular far-infrared sensors was studied.In order to meet the requirements of night-time driving observation range,the stitching technology based on binocular infrared images was studied,which was realized using scale invariant feature transform(SFIT)and random sample consensus(RANSAC).The experiments show the feasibility of the algorithm and the effect of the image stitching algorithm on expanding the range of view,and the security of pedestrian detection is further improved.