Research And Implementation Of Autonomous Driving Visual Perception Technology Based On Deep Learnin

With the rise of artificial intelligence technology,autonomous driving systems have become a new research focus in the field of automobiles.In autonomous driving systems,visual environment perception technology can provide precise positioning for vehicle control decisions.Visual perception technology needs to achieve accurate perception of vehicle targets and traffic signs,and obtain precise coordinate information.It also needs to achieve pixel-level visual perception of lane lines and road environments,and delineate lane lines and drivable areas.In this paper,the following aspects of research were conducted using deep learning:Firstly,addressing the issue of a large number of small targets in road traffic signs,this paper conducted research on existing object detection networks and used YOLOv5 s as the baseline model for object detection network construction.Based on YOLOv5 s,the ESE attention mechanism was introduced to simplify the CSP structure,and the DFM feature fusion module was proposed to optimize PANet for feature map fusion.Optimized K-means++ was used for reassignment of anchor boxes,and the loss function was optimized.The improved network was trained and validated on the TT100 K dataset,and the average precision m AP-0.5 for 45 classes of objects on TT100 K was achieved at 86.3%.Secondly,this paper constructed a multi-task network MTNet for autonomous driving visual perception assistance system,which achieved road vehicle detection,lane line segmentation,and drivable area delineation.The network encoder was constructed using an improved CSP structure,and the FEM feature enhancement module was proposed to improve lane line detection performance.GSConv and depth-wise separable convolution were introduced to reduce network computation.Multi-task loss was used for joint training of the multi-task network end-to-end.Experimental data on the BDD100 K dataset demonstrate that MTNet achieves a target detection m AP-0.5 of77.1%,an impressive drivable area m Io U of 97.8%,and a lane line detection Io U of27.5%.Finally,the two models proposed in this paper were deployed and tested using the YOLOv5s-dfm traffic sign small object detection algorithm on the embedded device Jetson Nano and the GPU platform.The multitask network MTNet was deployed and tested on a high-performance GPU.The experimental results demonstrate that both algorithms were able to achieve the experimental objectives,and exhibited satisfactory detection performance for various road conditions.