Research On Pedestrian Detection Algorithm And Security In Autonomous Driving

Nowadays,autonomous driving technology has gradually received extensive attention.At present,researches on autonomous driving technology are focused on mobilizing various functional modules,optimizing intelligent algorithms,and finally landing autonomous vehicles.Pedestrian detection technology as an important part of it,the scientific research workers and car manufacturers have gradually paid attention to it,and have proposed many excellent recognition algorithms.Pedestrians are one of the main targets in the road environment,so it is of great significance for the research of pedestrian detection algorithms and security.In this dissertation,the Tiny-Scale3-YOLOv3 pedestrian detection algorithm is proposed for the problem of missed detection and misdetection caused by the degree of darkness of the environment and the simple network structure of Tiny-YOLOv3;this dissertation proposes an AS-GAN pedestrian trajectory prediction algorithm for the problem that pedestrian interaction information cannot be fully utilized for trajectory prediction;this dissertation designs a pedestrian detection and trajectory prediction system to demonstrate the pedestrian location results obtained by Tiny-Scale3-YOLOv3,and the pedestrian trajectory predicted by AS-GAN.The main research work of this article consists of three parts,as follows:(1)In view of the requirement that the autonomous driving system can detect pedestrian targets in the vehicle’s surroundings in real time,the existing algorithm has the problems of missing detection due to different pedestrian poses and clothing,and the degree of darkness of the environment,as well as pedestrians as small targets.This dissertation presents the Tiny-Scale3-YOLOv3 pedestrian detection algorithm.This algorithm combines with the preprocessing of the image to adapt to different lighting levels,including gray processing,contrast processing,and the use of image resolution enhancement after local cropping.The model optimizes the Tiny-YOLOv3 network structure,and divides the collected image into two steps simultaneously: global detection and local detection,and the results of the two detections are fused to improve the accuracy of pedestrian detection.Experiments show that the accuracy of the method in Chapter 3 is higher than that of the comparison models,and it has good real-time performance.At the same time,the analysis of the loss of the algorithm has good stability and robustness,which can guarantee the security of the algorithm.(2)Aiming at the problem that social Generative Adversarial Networks(S-GAN)cannot fully utilize pedestrian interaction information,this dissertation presents a pedestrian trajectory prediction algorithm based on attention mechanism and social generation adversarial network(AS-GAN),which refers to the idea of Generative Adversarial Networks,retains the “social pool module” fusion method based on S-GAN to learn a global pooling result and enable LSTM to share information,and by combining the attention mechanism improves the time requirements of S-GAN to fuse information of pedestrians around,can fuse information at every moment,and selects the location information of pedestrians around,only fuses information that is useful to current pedestrians,improves the performance of the model,and then establish a pedestrian trajectory prediction model that can accurately depict pedestrian interaction patterns.The experiments prove that the prediction accuracy of the method in Chapter 4 is higher than that of the comparison model.At the same time,the analysis of the loss of the algorithm has good stability and robustness,which can guarantee the security of the algorithm.(3)A pedestrian detection and trajectory prediction system is designed,in which the function part of the pedestrian detection system can detect the position of the pedestrian in real time by selecting the picture or video file captured by the camera and give the confidence level,so that the autonomous driving system can respond promptly.In the function part of the pedestrian trajectory prediction system,select the existing data set,and the follow-up movement trajectory of the observation target can be given for the reference of the autonomous driving system.The movement trajectory is planned in advance to avoid pedestrians to prevent traffic accidents.