Research On Key Technologies Of Pedestrian Detection In Vehicle-Borne Image

As an important research task of computer vision,pedestrian detection is widely used in security,traffic control,automatic driving and virtual reality.In recent years,due to the rapid iterative update of image processing methods and graphics processing units,pedestrian detection technology has also made rapid development,in which pedestrian detection based on vehicle-borne image has also made a lot of research and application.It is found that millions of people die in road traffic every year in the world.Many studies show that pedestrian inspection on vehicle-borne image can effectively reduce the occurrence of traffic accidents.At the same time,the investigation of drivers also finds that they have a high acceptance of the pedestrian detection system based on vehicle-borne image.This shows that the research of pedestrian detection based on vehicle-borne image has great demand and market potential,so the development of pedestrian detection technology has become an important task.Pedestrian detection in vehicle-borne image refers to the use of various visual sensors to perceive the surrounding environment and road conditions on the moving vehicle,and make decisions through algorithms according to the perception information,so as to detect the pedestrians and their positions in front of the vehicle in real time and accurately,and then take appropriate preventive measures to help ensure the safety of pedestrians and drivers.Pedestrian detection is different from the general object detection.It needs higher accuracy to ensure the safety of people.At the same time,it also requires faster detection time to reduce traffic accidents.However,pedestrian detection based on vehicle-borne image is facing open traffic scene,which is easily affected by complex detection background,pedestrian scale imbalance and illumination,which brings great challenges to pedestrian location and classification in images.Therefore,pedestrian detection performance is still low.This paper focuses on the pedestrian detection challenges,such as complex detection background,multi-scale and insufficient illumination.The main research work and innovations of this paper are as follows:1)An anchor-free pedestrian detection method based on semantic segmentation is proposed.The method consists of two stages:region proposal generation and classification.In the stage of region proposal:firstly,an anchor-free region proposal generation network is designed,which can avoid the over parameter problem of artificial design in anchor-based candidate region generation method and reduce the miss detection rate of region proposal.Secondly,the semantic segmentation is used to guide and supervise pedestrian detection in the image,and the shared feature layer is generated with the anchor-free region proposal generation network,so as to enhance the pedestrian region features in the shared feature map and reduce the redundant features in the complex background region.For visible pedestrian dataset,a box-level semantic segmentation method is proposed to guide pedestrian detection;For far-infrared pedestrian data set,pixel-level labels are generated by traditional threshold segmentation for semantic segmentation to guide pedestrian detection.In the classification stage,deconvolution and context information are proposed to improve the discrimination ability of target classification network.Finally,the proposed algorithm is tested and compared on visible data set and far-infrared data set.The results show that segmentation can effectively suppress complex detection background,reduce false detection and missed detection.At the same time,the discrimination ability of classification network combined with deconvolution and context information is also improved.The final detection results show that the proposed pedestrian detection method with anchor-free on image segmentation can effectively suppress the complex background problems in pedestrian detection and ensure the accuracy of pedestrian detection.The final detection results show that the proposed anchor-free pedestrian detection method based on semantic segmentation method can effectively suppress the complex background in pedestrian detection and ensure the accuracy of pedestrian detection.2)A pedestrian detection method based on adaptive multi-scale feature fusion is proposed.Firstly,an adaptive channel feature fusion method is designed,which can learn the fusion weights according to the scale of pedestrians,so as to select different levels of features for fusion.Then,taking these fusion features as input,a multi-scale anchor-free(MSAF)network is designed to generate multi-scale region proposals.Finally,the proposed pedestrian detection method is evaluated on visible and far-infrared pedestrian detection dataset.Experiments show that the proposed pedestrian detection method can detect not only small-scale pedestrians,but also large-scale pedestrians.In general,the detection method proposed in this paper can solve the multi-scale problem well,so as to improve the detection performance.3)Attention fusion for one-stage multispectral pedestrian detection method is proposed.Firstly,a multispectral channel feature fusion method MCFF based on illumination attention mechanism is proposed to fuse the image features of different modes.Then,according to the starting stage of multispectral fusion,a variety of fusion frameworks are proposed to detect pedestrians based on YOLOv5.Finally,in the multispectral dataset,it is found that the proposed detection framework has the best fusion effect on the halfway fusion stage,and the performance of single-mode pedestrian detection is much lower than that of multispectral pedestrian detection in both day and night.Therefore,the pedestrian detection based on attention mechanism and multispectral fusion proposed in this paper can achieve good detection results in the case of insufficient illumination or night,and can solve the problems of insufficient illumination and night detection.