Font Size: a A A

Research On Lane Line Detection Algorithms Based On Deep Learning

Posted on:2020-11-11Degree:MasterType:Thesis
Country:ChinaCandidate:D F ZhangFull Text:PDF
GTID:2392330596484716Subject:Applied Statistics
Abstract/Summary:PDF Full Text Request
The correct perception of the environment for driverless cars is a necessary condition to ensure safe driving,so the correct detection of lane lines is very important.The traditional lane detection methods rely on manual extraction of features,when the environment change,the features also change,resulting in the algorithm robustness is weak.Deep learning image semantic segmentation technology via trains a large number of samples and learns features autonomously to achieve end-to-end pixel-level segmentation.It has strong robustness in complex driving environments such as illumination changes,shadows,nights,and corners etc.In this paper,the semantic segmentation algorithm based on deep learning is used to study the detection of lane,and the following research results are obtained:(1)For the problem that the traditional lane detection method the robustness is weak and the detection accuracy is low,in this paper,the deep learning semantic segmentation algorithm Deeplabv3+ is applied to lane detection.The detection accuracy of the algorithm under the five different data sets of Udacity dataset,Caltech dataset,Tesla dataset,Tusimple dataset and Gy dataset reached 83.74%,89.05%,92.56%,88.06% and 88.92%,respectively.They are higher than SegNet algorithm by 1.52%,26.58%,11.01%,5.86%,and 11.63%,respectively,and achieve better detection accuracy.(2)In this paper,these new convolution methods are proposed: row convolution and column convolution.Row convolution is to slice the features of ordinary convolution and then convolve between these slices;column convolution is to slice the features of ordinary convolution and then convolve between these slices.The experimental results show that the training error and verification error of the model with row convolution and column convolution are 0.0020 and 0.0020,respectively,while the training error and verification error of the comparison model containing only ordinary convolution are 0.0111 and 0.0098,respectively.And the models with row convolution and column convolution can accurately detect lane lines,while comparison models cannot detect any lane line information.Therefore,for image semantic segmentation,the proposed row convolution and column convolution have better segmentation effects.(3)For the problem of slow detection speed of the lane for Deeplabv3+ algorithm,A RaC_CNN lane detection method based on row convolution and column convolution is proposed.The detection speeds of the algorithm in the five different data sets of Udacity dataset,Caltech dataset,Tesla dataset,Tusimple dataset and Gy dataset are 5.92 frames per second,7.16 frames per second,6.13 frames per second,5.87 frames per second,6.07 frames per second,respectively.The detection speeds are 4.93 times,5.73 times,5.02 times,4.89 times,4.86 times of the Deeplabv3+ algorithm,respectively.The detection accuracy is 83.25%,82.54%,88.35%,89.73% and 89.17%,respectively.Obviously,the detection accuracy of RaC_CNN algorithm proposed in this paper is slightly lower than that of Deeplabv3+ algorithm,but the detection speed of RaC_CNN is much higher than that of Deeplabv3+ algorithm.The RaC_CNN method improves the detection speed while ensuring a certain accuracy rate,and achieves a good balance between speed and accuracy.
Keywords/Search Tags:Lane Detection, Semantic Segmentation, Deeplabv3+ algorithm, Row Convolution, Column Convolution
PDF Full Text Request
Related items