Research On Lane Line And Road Target Detection Based On Deep Learning

With the development of deep learning,autonomous driving technology has triggered a wave of artificial intelligence.Many traditional automotive companies,new forces in car manufacturing and high-tech companies have focused their attention on self-driving cars.As one of the key technologies,the detection of lane lines and road targets plays the most important role in the task of driving environment perception.Lane line detection is a prerequisite for the realization of advanced driving assistance systems,such as lane keeping assistance,lane departure warning,lane center assistance and other functions,as well as a necessary prerequisite for the realization of fully unmanned path planning and decision-making and vehicle motion control.How to accurately identify vehicles and pedestrians on the road in real time is a challenging and urgent problem faced by autonomous vehicles at present.Visual tasks based on the deep learning method has been beyond the traditional computer vision algorithms in various aspects.On the basis of summarizing existing achievements of image processing technology based on convolutional neural network,monocular vision and deep learning methods are used to carry out research on lane and road targets(such as vehicles,pedestrian)detection,which mainly includes the following aspects:1)The research status of lane detection technology and target detection is analyzed,and determines the key technology routes that need to be solved in the current research.In addition,we collect and produce data sets for algorithm validation and evaluation.2)Aiming at the task of lane identification detection,an algorithm based on instance segmentation is established.By analyzing the target structure of the lane,an improved lightweight Novel Bilateral Segmentation Network,Bisenet V2,is proposed for lane and background segmentation.In view of the various shapes of lane,the identification lines are processed into continuous solid lines to segment network learning features;In order to improve the detail attention of the network detail branch to the lane mark,the dilated convolution was used to increase the convolution receptive field,what’s more,3x3 depth separable convolution and 1x1 convolution are used to replace the original standard convolution to reduce the model Bilateral Network with Guided Aggregation for Real-time Semantic Segmentationattention mechanism.The network model’s learning center is shifted to channel feature graph with important features;In order to balance the unbalanced proportion between the lane and background,the weighted cross entropy is used as the loss function;Different from other post-processing methods with complex semantic segmentation,the maximum confidence coordinate point on the interest line is used as the candidate point for lane identification,and the least square algorithm is used to fit the segmentation results of cubic curve correction.3)Aiming at the problem of road target detection,through the analysis of the existing single-stage and two-stage target detection algorithms,the two-stage algorithm is complex,the model is large,and it is not suitable for direct deployment in the mobile terminal to achieve high real-time.Thus,a single stage target detection algorithm is determined.In addition to guaranteeing real-time detection,accurate detection effect should also be achieved for small targets.Improved a lightweight YOLO V4 algorithm suitable for direct deployment in driverless cars.Mobile V3 was used as the backbone network to extract features,the standard convolution in the decoding part of the algorithm is replaced by a deep separable convolution to reduce parameters and compress the model.At the same time,the Pyramid Pooling Module PPM is used to optimize the detection accuracy of small targets.In order to fit the input size of the network,boundary compensation was used to fill the data set into a square,and the resolution was changed again to input the network model for training and testing.Moreover,K-means algorithm is used to generate the priori box sizes required by different data sets.Finally,network detection performance is evaluated by comparing with other algorithms.4)The lane detection and road target detection algorithms are executed in parallel,after combining the two results,the joint detection of lane,vehicles and pedestrians is realized.This multi-task fusion detection method provides convenience for environment perception of autonomous driving vehicles.The size of the improved lane detection model is 12.4MB,and the target detection model is 41 MB,completely meet the deployment requirements of mobile terminal,and because it is two independent algorithms,the real-time detection is not reduced.In conclusion,the above research is to establish the accurate lane detection results,road target detection as a supplement,it is oriented to the larger goal of detecting the surrounding environment,such as vehicles and pedestrians,and integrates the two detection tasks.While realizing the road environment perception of autonomous vehicles,it effectively compresses the model,especially improves the accuracy of lane line detection and the real-time performance of multi-task and multi-target detection.