| With the rapid development of our country’s economy,the number of automobiles is increasing year by year.Increasingly congested roads make frequent traffic accidents,which not only cause huge economic losses,but also seriously threaten people’s lives.Therefore,driving safety issues and assisted driving have become hot spots in society.According to statistics,about 50% of traffic accidents are caused by vehicles deviating from the driving lane,so the automatic positioning of lanes has become an important task to assist driving,and lane line detection is the key among them.In practice,lane line detection is challenging.On the one hand,lane lines are often worn,blocked or even missing.Bad weather and complicated terrain have also greatly interfered with lane detection.On the other hand,the assisted driving system has high requirements for real-time performance,so how to improve the lane line detection speed under the condition of limited computing resources is also a problem that cannot be ignored.Therefore,how to design an accurate,fast and robust lane line detection algorithm is a challenging and valuable research topic.This paper proposes a lane line detection algorithm based on GAN,which can achieve accurate and fast lane line detection under roads with lane line wear and various external factors.The main work of this article is as follows:Aiming at the problem of the lack of robustness and real-time performance of the current lane line detection algorithms,this paper applies GAN to the field of lane line detection,and uses the style transfer network pix2 pix to convert the lane line detection task into an automatic generation learning from image to image task.The pix2 pix network can ignore the overall structure of the original image when generating the image,and focus on the extraction of detailed features,so it can ignore the interference of other external environmental factors on the lane line while extracting the features of the lane line,improving the lane line Detection robustness.Applying GAN network to detect lane lines does not require any prior knowledge and subsequent operations.The algorithm structure is simple,so the detection time and running space have been greatly improved.In order to verify the effectiveness of the algorithm,tests were performed on the PreScan simulation platform and TuSimple dataset: the road detection situation under various external environmental interferences was compared and analyzed and quantitatively evaluated to verify the robustness of the algorithm;The average detection time of each image is 0.003 s,which is two orders of magnitude lower than the 0.163 s of the comparative experiment LaneNet,and the memory occupation is 801 M,which is 88% less than LaneNet’s 6485 M,which proves the algorithm’s obvious advantages in detection time and cost.Aiming at the problem of missing and wrong detection of lane lines,a fw-pix2 pix lane line detection algorithm is proposed.Hierarchical feature matching and Wasserstein distance are introduced into the algorithm to further improve the detection accuracy of the algorithm.In order to better constrain the detection quality of lane lines,feature matching is used to discriminate the generated images in each layer of the discriminator.By extracting the features represented by the layers of the discriminator,the similarity between the generated image and the original data is restricted,and the details of the generated image are guaranteed layer by layer,so that more complete lane line features can be extracted.At the same time,in order to solve the problem of disappearing gradients during training,the Wasserstein distance was introduced to measure the distance between the two distributions of the generated image and the original data,which further constrained the quality of the lane line detection image generation and reduced the missed and false line detection.In order to verify the effectiveness of the algorithm,the TuSimple dataset was used to test the fw-pix2 pix lane line algorithm in this paper.The Accuracy index of the algorithm is 97.9%,the F1 score is 79.7%,the average detection time of each image is 0.003 s,and the memory occupation is 801 M,which shows that the fw-pix2 pix algorithm maintains the advantages of detection time and cost and further improves the detection accuracy.The algorithm in this paper has obvious advantages in detection time and cost while having good robustness,which ilustrates the feasibility and practicability of the algorithm in real application. |
