Research On Autonomous Driving Target Detection Algorithm Based On Deep Learnin

As an important branch in the field of autonomous driving,target detection maintains road traffic safety by detecting environmental targets around the road in real time.Traditional selfdriving target detection extracts features by manual design,which has high computational overhead and poor robustness,making it difficult to meet the needs of actual self-driving target detection.While,deep learning-based target detection has stronger feature extraction capability by automatically extracting input image feature information.Therefore,it is extremely important to research deep learning-based target detection technology for autonomous driving.Aiming at the problem of small scale and low resolution of traffic signs,a YOLOv5 s traffic sign detection network fusing Transformer and improved PANet(Path Aggregation Network)is proposed.With fusing Transformer coding module,the original PANet structure is improved,and the corresponding prediction candidate frames are derived by K-Means algorithm,meanwhile,the cross-entropy loss function of the original network is improved.Experimental results demonstrate that the improved YOLOv5 s network improves the detection accuracy and recall rate by 2.2% and 0.7%,while reducing the number of parameters and computational complexity by 25.8% and 10.1%,as compared to the original network.Aiming at the problems of severe occlusion,diverse weather changes and complex backgrounds of targets such as people,vehicles and traffic signs in actual autonomous driving scenes,a complex driving scene target detection network incorporating 3D attention and weighted frames is proposed.The parameter-free Sim AM attention mechanism is fused and the non-maximum suppression of the original network is improved.Experimental results demonstrate that the improved YOLOv5 s network improves detection accuracy and recall by4.3% and 4.9%,while decreasing the number of parameters and computational complexity by16.7% and 6.3%,as compared to the original network.Aiming at the difficulty of deploying practical self-driving target detection algorithms,an embedded platform-based approach for self-driving target detection is used.The above two improved YOLOv5 s networks are deployed on the NVIDIA Jetson AGX Xavier embedded platform and further optimized using Tensor RT inference.Experimental results demonstrate that the above two improved YOLOv5 s with Tensor RT inference optimization improve FPS by13.9% and 14.4%,while compared to the Xavier-only platform,and are more suitable for deployment on resource-constrained in-vehicle platforms.