Research On Multi-Object Perception Algorithm For Autonomous Vehicles Based On Deep Learning

With the continuous development of artificial intelligence and computer vision in China,unmanned driving has become a research hotspot in the field of artificial intelligence.The perception system of unmanned vehicles,as a core component,is the most important module to ensure the traffic safety of unmanned vehicles.In view of the problems of traditional vehicle perception systems,such as redundant multiple models,poor accuracy,and poor realtime performance,this article combines deep learning and multi-task learning technologies to conduct research on the perception algorithms of unmanned driving vehicles.The main research contents of this thesis are as follows:The shortcomings and challenges in the practical application of current vehicle detection models were analyzed,and a lightweight vehicle detection algorithm based on linear attention was proposed.The attention module in the model is improved to accelerate the model’s inference speed.The Mosaic data augmentation structure in the preprocessing module was also improved to increase the data for blurred and small vehicles,making the algorithm more robust.The EIo U loss function was used to optimize the boundary box regression effect of the algorithm,improving the accuracy of vehicle detection.Comparative experiments were conducted on the BDD100 K and KITTI datasets to verify the improvement and performance of the vehicle detection algorithm.Proposed an improved TransTrack multi-object tracking algorithm for unmanned driving vehicles.Based on the Trans Track multi-object tracking algorithm,two improvements are made: introducing the Mish activation function in the Res Net feature extraction network to alleviate the gradient disappearance problem caused by the high proportion of small targets and improve the performance of the detector;using the improved data association algorithm to match all detection results to improve the tracking of occluded targets and reduce the loss of targets.A panorama perception algorithm for autonomous vehicles based on YOLO and multitask learning is proposed.The algorithm combines traffic target detection,drivable area segmentation,and lane line segmentation tasks.Through dynamic attention,the algorithm improves the model’s performance in three aspects: scale perception,spatial perception,and task perception,which enhances the accuracy of the model in the three aforementioned tasks.The study also adjusts the multi-task loss function and introduces Focal Loss to address the challenge of imbalanced data,further improving the accuracy of the vehicle perception algorithm for autonomous vehicles.