Research On Key Technologies Of Visual Perception For Advanced Driving Assistance Systems Based On Deep Learning

In recent years,advanced driving assistance systems(ADAS)has received widespread attention and rapid development.1)The development of information technology in the human society and people's continuous new demands for travel experience.2)The collision between the artificial intelligence technology,which is an important part of computers,and the automobile manufacturing industry,which is at the top of the industrial manufacturing pyramid.ADAS is a comprehensive system integrating terminal computing,environment perception,decision planning and other modules.It is deployed on vehicles to provide drivers with forward collision reminders,traffic sign reminders,road maintenance and other functions to improve the safety of drivers during driving.It can effectively reduce the incidence of vehicle accidents.With the development of deep learning in the image field,ADAS based on visual perception has been favored by academia and industry due to its low hardware cost and excellent performance.This article focuses on lane detection,pedestrian detection and traffic sign recognition in ADAS machine vision.The specific work is as follows:(1)Aiming at the problem of category imbalance and interference in lane detection,a lane detection model F-FCN is proposed.A pixel-level UCAR lane dataset is established.Based on the fully convolutional network(FCN),a fusion structure based on feature splicing is constructed to alleviate the interference problem.Focal loss is used to alleviate the category imbalance problem in lane detection.The genetic algorithm is used to search for the optimal parameter combination in Focal loss.The experimental results on the UCAR lane dataset show that F-FCN can improve the detection accuracy of the basic FCN from 83.6%to 88.9%(mPA).(2)Aiming at the problems of small targets,occlusion and background interference in pedestrian detection,pedestrian detection models E-SSD and Mask-SSD are proposed.A UCAR pedestrian dataset labeled with bounding boxes is built.E-SSD uses a one-stage detection framework as the basic network.A feature fusion structure based on deconvolution to alleviate the problem of small targets is built.An attention mechanism based on channel attention and spatial attention to alleviate occlusion is built.Mask-SSD also uses a one-stage detection framework as the basic network.A semantic segmentation branch based on dilated convolution is constructed.It generates a mask based on semantic segmentation features and provides a target search area for the detection process.The experimental results on the UCAR pedestrian and Caltech pedestrian datasets show the excellent performance of E-SSD and Mask-SSD in detection speed and detection accuracy.The miss rate on Caltech pedestrian data set is 50.08%and 48.20%respectively,and the detection speed is 0.11 s/frame and 0.13 s/frame.(3)Aiming at the problem of small targets and similar category interference in traffic sign recognition,a traffic sign recognition model Dense-RefineDet is proposed.A UCAR traffic sign dataset labeled by bounding box is built.DenseRefineDet uses a one-stage detection framework as the basic network.A dense connection-based feature transfer structure is constructed to alleviate the similar category interference problem.An anchor generation method suitable for traffic signs is designed to alleviate the small target problem.The experimental results on the UCAR traffic sign and Tsinghua-Tencent 100K datasets show the excellent performance of Dense-RefineDet in detection speed and detection accuracy.On the Tsinghua Tencent 100k dataset,the recall of small-size targets is 84.3%and the precision is 83.9%.(4)Considering the future development direction of ADAS vision algorithm,a two-branch weakly supervised object detection model is proposed by the transition from supervised algorithm to weakly-supervised algorithm.The twobranch model structure includes a detection branch and a self-attention branch.The dual-branch structure is used to transform the weakly supervised detection problem into a supervised detection problem.The two branches use each other's segmentation mask as the pseudo label.The experimental results on the Stanford cars,CUB-200-2011 and VOC2007 datasets show the effectiveness of the twobranch weakly supervised object detection model.On the VOC2007 dataset,the pointing accuracy is 82.20%,and the result of Corloc is 30.65%.