Research On Lane Line Detection By Deep Learning In Night Time Environment

In recent years,Advanced Driver Assistance Systems have been developing rapidly.As one of the key aspects of Advanced Driver Assistance Systems,all-weather lane line detection methods have attracted extensive attention from major research institutions.The detection method based on visual image features dominates the existing lane line detection algorithms and is one of the future development directions.At present,the research method of lane line detection based on neural network is more mature under good lighting conditions.However,given the complexity of illumination in nighttime situations,detection algorithms that have high detection performance under good lighting conditions are not well suited for nighttime environments.In this paper,from the perspective of improving the performance and real-time performance of lane line algorithm detection,weak illumination image enhancement method,lane line detection method based on image enhancement and knowledge distillation,The main research contents of this paper are as follows:(1)The existing low light image enhancement algorithm has the problems of color distortion and blurred image details in the images obtained after processing the low light images.Therefore,an image enhancement algorithm based on Retinex theory is proposed in this paper.For the problem of color distortion caused by algorithm defects,the color correction branch network is embedded in the enhancement network part of Retinex Net network in a parallel way to reduce color deviation;for the problem of blurred details,the Convolutional Block Attention Module is firstly embedded in the illumination enhancement network to perform at the same time,bilinear interpolation is used in the upsampling part of the network to enrich the information of the network.The experimental results on the low-light dataset show that: compared with the existing image enhancement algorithms,the output results of this paper have less color distortion and more outstanding details when the algorithm enhances the low light.However,the algorithm is applied to nighttime lane line images,and the obtained enhanced images contain more noise and interference information,resulting in unsatisfactory enhancement.(2)To address the problem that nighttime images contain a large amount of noise and the contrast between the lane lines and the road surface causes the lane line features are not easily detected,a lane line detection method ZERO-UFAST based on unsupervised image enhancement network is proposed.firstly,in order to take into account the enhancement of luminance and the suppression of noise,the input initialization of the network is changed to three times different scales of expansion convolution merged on the channel,constituting a lane line detection method with light details more feature information,and reduce the number of output channels of the light mapping curve network to suppress the noise in the enhanced part of the image;considering that the lane lines at night are not obviously featured in the road background,the two-stage images of enhanced light,light mapping and then enhanced are fused to improve the contrast of the road area.The performance analysis test is conducted on the CULane dataset,and the results show that the detection performance of the algorithms of different backbone networks in the nighttime environment are more significantly improved,and the nighttime detection F1 reaches 67.0 when Res Net50 is used as the backbone network.(3)To address the problem that the ZERO-UFAST detection algorithm with Res Net50 as the backbone network has high complexity leading to poor real-time performance,a lane line detection algorithm based on knowledge distillation is proposed.First,the network architecture of lane line detection based on knowledge distillation is proposed in the training section: the ZERO-UFAST network with Res Net50 as the backbone network is set as the teacher branch network,and the ZERO-UFAST network with Res Net18 as the backbone network is set as the student branch network.Between the auxiliary segmentation outputs of the teacher network and the student network,the KL scatter loss function was used to enhance the lane line feature extraction capability of the student network backbone;between the anchor detection outputs,the mean square error loss was used to bring the student branch network outputs closer to the teacher network;thus the student network has detection performance close to that of the teacher network.In the testing phase,the student branch network is used for lane line detection,and the number of parameters is reduced to improve the real-time performance of the algorithm.Performance analysis was performed on the CULane dataset,and the algorithm in dissertation has 17.5%of the detection time and 1.3 decrease in nighttime lane line detection performance F1 compared to the ZERO-UFAST lane line detection algorithm with Res Net50 as the backbone network.