Implementation Of Diggings Drivable Area Segmentation And Vehicle Detection Technology Based On Deep Learning

With the development of deep learning technology,more and more applications have been realized.Unmanned driving is a product of the new generation of artificial intelligence waves,and it is also a hot topic of current research.It can not only reduces human costs,but also greatly reduces accident rate.There are more and more application scenarios of driverless technology,such as cargo transportation,passenger taxis,etc.This article designs a model of driving area segmentation and target detection in the unmanned driving scene of the mining area,and it is deploys to the embedded side for performance testing in the actual scene.The map-side semantic segmentation,mining card-side semantic segmentation and target detection tasks of the mining scene are completed.The research contents of this article are as follows:Firstly,the research background of this topic is presented,and related research methods at home and abroad are introduced.The current application and development context of deep learning are sorted out,and the development process of deep learning is explained in conjunction with the research content of this article.Based on the current research status,the main research content of this topic is analyzed,and the technical requirements of related fields are briefly explained.Secondly,the environment of the mining area is introduced and the challenges of it are pointed out.Some basic algorithms for image processing to solve the environmental needs of the mining area are proposed based on the environmental characteristics of it.In addition,the technical solutions required in the special environment of unmanned driving in the mining area are explained.For the tasks in the mining area environment,the semantic segmentation and target detection image processing technologies were selected to highlight the spatial characteristics of the convolutional neural network structure and the temporal characteristics of the recurrent neural network.Thirdly,the basic network structure required for the segmentation of the drivable area on the map side and the car side is analyzed,and the model suitable for the mining area scenario is selected for iterative optimization.The map end requires high accuracy,and the truck-side needs high speed while ensuring accuracy requirements.The map-side segmentation task is completed at the ground station in the mining area.Next,the map information is transmitted back to the car-side,and the car-side considers its own segmentation result and the map-side information to realize the road segmentation task of autonomous driving.Finally,the vehicle detection task at the car end is completed,and the vehicle detection model at the mine card end is deployed to the Xavier embedded end.A lightweight network structure is selected to implement the detection task,because it is completed on the vehicle side and needs to ensure real-time performance.On the basis of the basic network,the convolution strategy and loss function are optimized,and the optimized network model is deployed on a high-performance Xavier board.