| The rapid development of computer vision technology has led to a significant change in intelligent vehicle perception systems.Environment perception systems can obtain real-time information about the vehicle’s surroundings,which is of great significance to ensure the driving safety of the vehicle.In complex traffic road scenes,the pedestrians captured by vehicle-mounted cameras are usually presented in the form of medium and small objects in the images and the phenomenon of missed detection and false detection often occurs,Therefore,this thesis proposes a pedestrian detection and tracking method for intelligent vehicles.It improves the performance of pedestrian detection and tracking while satisfying real-time requirements,especially significantly improving the detection and tracking of small-scale pedestrian at long distances,and providing rich visual information for subsequent multi-sensor fusion of intelligent vehicles.Firstly,a YOLOv5s-based pedestrian detection method(SD-YOLOv5)is proposed to address the problems of false detection,missed detection,and poor detection of small objects in pedestrian detection.Based on YOLOv5 s,the network model is improved in lightweight,and the Ghost Bottle Neck module is used instead of the Bottle Neck CSP module.In order to reduce the error between the ground truth box and the predicted box,Alpha-Io U is introduced as the localization loss function.Slicing aided hyper inference(SAHI)is applied to improve the inference part to improve pedestrian detection accuracy.The experimental results show that compared with the benchmark YOLOv5 s on the BDD100 K dataset,SD-YOLOv5 reduces the computation amount of the network model by 5.5 GFLOPs,reduces the number of parameters by 2.2M,increases the AP by 4.8%,and increases the APs of small-scale pedestrians by 2.1%.While meeting the real-time requirements,the detection accuracy of pedestrians is improved,especially the detection effect of long-distance small-scale pedestrians is significantly improved.Secondly,a Deep Sort-based object tracking method(LW-Deep Sort)is proposed to better improve the efficiency of pedestrian tracking.Based on Deep Sort,a lightweight OSNet network is used to replace the original network model to improve the efficiency of object appearance feature extraction.The experimental results show that compared with the benchmark Deep Sort on the MOT16 dataset,LW-Deep Sort improves MOTA by 1.6 and the speed by 5.3HZ,which significantly improves the performance of the tracker.Finally,Tracking test of SD-YOLOv5 pedestrian detection method with LW-Deep Sort tracking method on video data of complex lighting scenes on real traffic roads.Tensor RT is used to optimize and accelerate the pedestrian detection and tracking method on the intelligent vehicle platform,and the pedestrian detection and tracking test of the campus traffic road scene are completed.The inference frame rate is maintained at 44 FPS,which meets the performance requirements of the intelligent vehicle environment perception system. |
