Research On Visual Perception Algorithm Of Complex Driving Environment Based On Deep Learning

Traffic environment perception technology is an important part of intelligent driving system,and it is the basis of traffic control and decision-making.Real-time and accurate target detection and identification technology provides guarantee for the safety and reliability of intelligent vehicles.Compared with lidar sensors,the environment perception technology based on machine vision has the advantages of low cost and rich perception information,which is widely used in the fields of image processing and target detection,and provides a new idea for the environment perception technology of intelligent cars.However,the changes of lighting,weather and road type are easy to affect the visual image,resulting in the lack of accuracy and real-time performance of the algorithm.Therefore,intelligent vehicles cannot realize accurate perception of different traffic environments through a single algorithm.This article selects the complex environment of the intelligent vehicle driving environment perception system as the research object,in view of the vehicle driving in the process of road driving,traffic signs and pedestrians in the detection and identification of key elements such as poor lighting,multiple points of view,multi-scale change problems such as conduct in-depth research,to realize the accurate perception of intelligent vehicle in complex environment.The main research contents and achievements of this paper are as follows:(1)Accurate detection of roadable areas is the most basic condition for intelligent vehicles to travel safely on urban roads.This paper proposes a U-Net-based detection network of RGB-D roadable areas(RGB-D-U-CRF)aiming at the problem that the existing algorithm has a poor ability to detect bad lighting conditions such as strong light,shadows,reflections and complex road types.Firstly,the dual-branch and post-fusion UNet network structure is used to extract the features of RGB image and depth map respectively to obtain the rough classification of drivable areas.Secondly,the RGB-D condition is constructed to optimize the predicted region boundary and realize the fine region detection.Compared with the representative algorithms in this field,the results show that the proposed algorithm has higher pixel detection accuracy,intersection ratio and better detection efficiency under bad care.(2)To solve the problem of insufficient detection accuracy of existing SSD and YOLO single-stage detectors,this paper proposes a weighted dense connection-based compression network(CWDC-YOLO)for traffic sign detection.On the basis of YOLOv3,a dynamic weighted dense connection block and a convolutional attention module are introduced to improve the expression of image information through the comprehensive utilization of feature graphs with different weights on the premise of not increasing the amount of network computation,and the idea of Mobile Net compressed sensing is adopted to further improve the efficiency of the algorithm.The verification of GTSDB data set shows that the traffic sign detection efficiency of CWDC-YOLO is 29 fps,which can meet the real-time detection requirements.Compared with YOLOv3,the accuracy can be improved by 6.21%,reaching 90.14%.Moreover,for the self-built data set of CTSDB,it also has good precision rate and recall rate.(3)To solve the problem of traffic sign recognition in the process of driving,this paper proposes a traffic sign recognition network(ETE-CSRNet)based on sparse coding end-to-end learning.In order to retain the position and parameter relationship among the image features in the training process,this paper uses Caps Net to extract the image features,and in the process of dynamic routing transmission,adopts the method of vector length and direction joint discrimination to improve the network's ability to identify samples with similar shapes.Secondly,a novel end-to-end sparse coding structure is constructed.This structure transforms the sparse coding minimization solution process into a multi-level recursive neural network layer,and finally realizes end-to-end training by combining with the multi-scale spatial pyramid pooling layer,so that the network can obtain more representational sparse coding.In this paper,the effectiveness of the proposed network is verified on the GTSRB data set,and the results show that,compared with various advanced algorithms,the ETE-CSRNet can obtain the highest recognition accuracy,and the performance of the unsupervised sparse coding network using convolution feature can be improved by 4.61%,and the traffic signs with the angle of deviation inside can still be effectively recognized.It also has good generalization ability for CTSRB self-built data set.(4)In the complex driving environment,the scale distribution of pedestrians is very wide,and the small and medium scale pedestrians account for a large proportion.To solve this problem,this paper proposes a multiscale pedestrian detection algorithm(MS-Faster RCNN).First of all,according to suggestion regional location inaccuracy problem,by combining with the characteristics of pedestrian scale distribution network proposed multi-scale regions(MS-RPN),and receptive field size as constraints on different convolution layer set size is more in line with the pedestrian detection of sliding window,the second for multilayer pool depth characteristics graph information loss problem,by building detection on the basis of characteristics of context fusion subnet to enrich expression image characteristics.The Caltech data set experiment shows that compared with the general RPN network,the recall rate of MS-RPN for mesoscale and small-scale pedestrian detection is increased by 10.8% and 4.8%,respectively,with obvious improvement effect,and the overall recall rate is increased by 3.2%.Compared with other pedestrian detection algorithms,MS-Faster RCNN also obtained the optimal detection result,the logarithmic average miss rate of Miss rate-FPPI curve was 9.26%,and it also had good detection performance for ETH and CPDB self-built data sets.Through the above research,a more accurate and rapid perception algorithm is provided for the deep learning-based intelligent vehicle visual perception system,which has certain theoretical and application value for ensuring the safety and reliability of intelligent vehicles in complex driving environment.