Key Technologies Of Intelligent Detection In Self-Driving Based On Deep Learning

With the popularization of in-vehicle intelligent equipments and the advancement of vehicle-road coordination systems,self-driving has received more and more attention.Intelligent detection is an important perception part of the self-driving system.It helps the self-driving intelligent terminal to plan driving by detecting the position of objects on the road.Fast and accurate intelligent detection can greatly improve the safety of selfdriving.Therefore,the research and application of self-driving intelligent detection are of great significance.This paper divides self-driving intelligent detection into object detection and lane detection.Object detection aims to detect objects such as vehicles,pedestrians,etc.Lane detection is responsible for detecting lanes.Intelligent detection methods based on deep learning have become the mainstream due to their powerful learning capabilities and the large-scale data support.Different from general scenes,detection methods in self-driving scene require not only higher accuracy,but also stronger real-time performance,to assist the self-driving terminal quickly and correctly handle various emergencies.Therefore,this research aims to design a fast and accurate self-driving intelligent detection system,providing real-time and correct road information for self-driving intelligent terminals.The main contributions of this paper are as follows: First,we research the network structure design and feature representation in object detection methods.In the network structure design part,different feature extraction operations and network structures are designed for three sub-networks,i.e.,backbone network,neck network,and head network.The network structure is universal and can greatly improve the accuracy of existing object detection methods.In the spatial and scale feature representation part,existing real-time object detection methods have simple network structures and strong real-time performance,but the accuracy of them is low.Therefore,we design an adaptive feature aggregation manner to enhance the spatial and scale feature representation ability,achieving a significant increase in accuracy with a small calculation overhead;second,we research the efficient feature extraction and feature information enhancement in lane detection.Based on Anchor-base framework and graph,we implement an efficient local and global extraction feature manner,which greatly enhances the discriminating and perception ability of the feature.To address the problems of information loss and insufficient communication in the processing of lane features,we introduce frequency domain learning into lane detection.It utilizes discrete cosine transform(DCT)to enhance the diversity of features,improving the detection accuracy;third,we combine the object detection method and the lane detection method into a real-time,accurate and complete self-driving intelligent detection system,which can detect road objects(such as pedestrians and vehicles),as well as lanes in self-driving scenes.It can provide fast and accurate road perception information for the self-driving intelligent terminal,and enhance the safety of self-driving.