Multi-Lane Line Detection And Research Based On Convolutional Neural Network

Autonomous driving can not only greatly improve traffic safety and road congestion problems,but also promote the transformation and upgrading of the automobile industry and boost economic development.Multi-lane line detection is one of the necessary functions of autonomous vehicles,which can help drivers better grasp the overall road information,and plays a vital role in the path planning and decision-making process of autonomous vehicles.The lane line detection method based on deep learning trains a large amount of sample data in advance,and obtains the final lane line model parameters through network learning.It shows a more powerful effect in feature extraction and greatly improves the performance of lane line detection.This paper mainly studies multi-lane line detection based on convolutional neural network.The main contents are as follows:In order to accurately segment the lane lines and improve the inference speed of the network model,this paper proposes a lane line segmentation network based on bilateral Lanenet.The shallow spatial detail information of the lane line is extracted through the detail branch,the deep semantic information of the lane line is extracted through the semantic branch,and a bilateral aggregation module is designed to fuse the characteristic information of the two.It has a good segmentation effect on the Tusimple data set,with an average intersection ratio(m IOU)of 0.728.For the lane line segmentation results of bilateral Lanenet,this paper improves the traditional RANSAC algorithm in the quadratic curve fitting.The three points initially selected are constrained,and new model evaluation criteria are used to determine the optimal model.Finally,when a better model is obtained,the points that do not meet the threshold are eliminated,which can accurately fit the lane line on the Tusimple data set.For lane line detection in complex road environments,a lane line detection model based on TF-Res Net18 is built.Before inputting the picture,the data set was preprocessed by data filtering and data enhancement.The lane line feature extraction adopts the TFRes Net18 network structure which integrates the improved Res Net18 network and the Transformer module.First,the Res Net18 network is improved,the number of channels of each module is reduced,and the detailed characteristics of the lane lines are initially obtained.Then combined with the Transformer codec module structure,in the environment of severely occluded lanes and strong illumination,the self-attention layer can effectively learn the contextual information of lanes.Finally,the FNNs module is used to directly predict the model parameters.Its accuracy on the Tusimple data set is96.28%,and the false detection rate FP value is 0.0282.At the same time,the F-meature value on the CULane data set after manual screening is 85.2,which verifies the effectiveness of the TF-Res Net18 network in this paper in multi-lane line detection.