Research On Autonomous Driving Road Scene Understanding Algorithm Based On Deep Learning

With the continuous development of deep learning and computing equipment,complex models can run in parallel computing equipment in real time,and use artificial intelligence technology to move from theoretical research to practical application.Scene understanding is one of the main areas of computer vision technology.It can decompose the elements of road traffic by capturing data in autonomous driving technology.It is one of the prerequisites for vehicles to make route planning and decision-making.For different traffic elements,structured decomposition is necessary to adapt to different detection methods.In autonomous driving scenarios,the real-time nature,sensitivity,rapid interpretation ability,and accuracy of the vision system are the guarantee of the autonomous driving system.How to balance speed and accuracy,and the edge computing equipment with reasonable algorithms,has great practical significance for the actual landing.In this paper,aiming at semantic problems in road scenes,by structurally decomposing traffic elements,using target detection and semantic segmentation algorithms,gradually through more refined recognition,to achieve a more refined understanding of the spatial location distribution between roads and their semantic information.The main research results and related summaries are as follows:1.From the overall feature to the captured local relationship is one of the keys to improve the accuracy of semantic segmentation.Here,a perceptual field module is proposed by using spatial cavity convolution and residual ideas,and gradually analyzes the image through comprehensive perceptual field coverage.The semantic relationship between the two,combined with the dense layer jump connection mode of the UNet semantic segmentation network,can achieve a reasoning speed of 61 FPS.2.Aiming at the positioning and recognition capabilities of target detection,a target detection network that combines depth separable convolution,spatial pyramid hole convolution,and dense layer skipping is proposed.The input image is 608 on the Nav Info Autonomous Driving Data Set.Under 608×608 resolution,it can achieve 44% mAP and12.5FPS,shorten YOLOv3 and Cascade R-CNN,which proves that its speed and accuracy can well adapt to edge computing performance.3.Aiming at the robustness of semantic segmentation,a post-processing algorithm based on morphology and maximum connected domain is proposed here to improve the elimination of error regions,holes,glitches,etc.in the segmentation results.Experiments prove that it can segment the results of semantics.There is a small improvement,which can effectively increase the stability of semantic segmentation.4.In order to improve the reasoning ability of the road scene understanding algorithm in edge computing,the network is deployed on the NVIDIA Jetson Xavier NX embedded platform through model simplification.The reasoning speed of the network for target detection is 6FPS,and the reasoning speed of the semantic segmentation network is 12 FPS.In summary,here we combine image processing,object detection,semantic segmentation,use the combination of different levels of scene understanding,and use deep learning to abstract features through images to express richer traffic information.Research,from theoretical derivation to model design,from algorithm verification to model deployment,has certain theoretical significance and practical value.