Research On The Detection Method Of Road Driving Area Based On Deep Learning

As one of the research directions of intelligent vehicle,the road environment recognition technology based on machine vision has become a research hotspot.Firstly,the complexity of the road environment lead to the unsatisfactory detection effect of the driving area.Secondly,intelligent driving requires high accuracy and real-time for road environment recognition algorithm.The robustness of the algorithm is poor,and the application conditions of the algorithm are harsh.Intelligent driving environment perception is the key foundation of vehicle decision-making and control.How to accurately and real-time identify the driving area of the road environment has become the top priority of this work,which is a major challenge in the field of intelligent vehicles.The driving environment was detected by Deep Learning image semantic segmentation algorithm based on the detection of driving area of intelligent vehicle road.The traditional lane line detection method has certain effect for simple road environment,but the detection effect for complex road environment was very poor.It is difficult for the algorithm to meet the requirements of automatic driving due to the false detection and missed detection.In view of this situation,the semantic segmentation algorithm based on convolution neural network was used to detect the driving area of the road.The standard model structure of Segnet network was improved to achieve the goal of semantic segmentation of driving environment.Firstly,the data set was acquired and annotated.The experimental data set came from two parts: One was used the existing open data set,and the other was used the labelme software to manually annotate the data set to obtain the test data set.Secondly,the Segnet semantic segmentation network model was constructed.Inspired by the U-net network,the parallel jump structure was used to connect the feature map of the encoding structure and the feature map of the decoding structure,which can make use of multi-layer information in classification and prediction.Then,the concept structure was used to replace the part of the network volume layerincreased the width of the network,reduced the number of network parameters.Finally,Adam algorithm was used to optimize the network training,and the road semantic segmentation network model was obtained.After the model was trained,a comparative test was carried out on the Selfmade data set and Camvid and cityscapes data set.The test results show that compared with the traditional lane line detection method,Deep Learning semantic segmentation algorithm has a good detection effect in various road condition detection environments,and the generalization ability of the model was strong,it can meet the requirements of real-time semantic segmentation of driving environment.