Research On Pedestrian Detection And Segmentation Of Vehicle Far Infrared Image

While the industrial revolution represented by autonomous driving systems is moving forward,it is also facing huge challenges.Especially on safety issues,how to make the system quickly and accurately find pedestrians in driving areas is a key issue.The use of computer vision technology to detect and locate pedestrians is becoming increasingly mature.At the same time,far-infrared technology is also shining in the field of image processing.Therefore,the study of using computer vision technology and far-infrared technology to detect and segment pedestrians is very meaningful for autonomous driving systems.The main work of this article includes the following:(1)The first is the collection and production of far-infrared images.This article analyzes the importance of far-infrared imaging for automated driving systems,explains how to obtain far-infrared image datasets and effectively label and preprocess pedestrian targets.On this basis,a low-resolution far-infrared dataset DLMU_FIR2020 is obtained.(2)The second is pedestrian detection in far-infrared images.In this paper,the single-stage object detection YOLO algorithm is selected as the basic framework.On this basis,an improved far-infrared pedestrian detection method is proposed.Compared with the original algorithm,the AP of the DLMU_FIR2020 and FLIR datasets are improved by 1.5%and 0.9%,respectively.(3)Finally,this article describes the segmentation of pedestrian objects in far-infrared images.This article is divided into two routes.The first is the traditional segmentation method that combines the object detection bounding box.Both methods achieved more than 60%on IoU in 30 random tests of DLMU FIR2020 and FLIR datasets.In the second route,a far-infrared pedestrian segmentation model based on deep learning is proposed.The IoU of the model in the KAIST pedestrian segmentation test set reached 46.79%,and the IoU in the 30 random tests of the DLMU_FIR2020 and FLIR datasets using transfer learning methods reached 36.57%and 66.24%respectively.In summary,the low-resolution far-infrared data set produced in this paper enriches the current research data,and the proposed vehicle-based far-infrared pedestrian detection and segmentation algorithm can effectively identify the pedestrians in actual driving scenarios.