Study On Lane Mark Detection Under Complex Environments

In recent years,autonomous driving has been a hot research topic in academia and industry.As a prevalent traffic sign in the road environment,lane marks play a crucial role in vehicle positioning and navigation.In real-world scenarios,lane marks vary in number and have a variety of appearance,which would suffer from wear,fading or stains.External environment such as bad weather,low light,and serious occlusion can also affect the visibility of lane marks.Therefore,lane mark detection must have strong robustness while maintaining high accuracy,which poses a great challenge for research in this area.Traditional image processing approaches can only achieve good lane mark detection under specific conditions.They are not suitable for practical applications.With the emergence of deep neural networks and large-scale datasets,lane mark detection based on deep learning has been significantly improved,but there are still some problems in complex road scenarios to be further solved.In order to obtain better lane mark detection performance for the practical requirements,this paper carry out an in-depth study of lane mark detection techniques in complex scenarios.The main work and contributions of this thesis are summarized as follows:1.To address the effects of road slope and camera pose changes on lane mark fitting,this thesis proposes a homography prediction network,HP-Net,which can adaptively learn to estimate the projection parameters of images.For whether the camera is pre-calibrated or not,two different homography projection matrix generation methods are designed.Using the prior that multiple lanes are parallel to each other,a label-sharing approach is adopted to train HP-Net without any extra data annotation.Experiments show that the accuracy of lane mark detection can be effectively improved by fitting the detected candidate lane points in a suitable top view space.2.For complex situations such as severe lane mark occlusion,wear,and stains,this thesis proposes a lane mark detection method based on multi-frame alignment and spatio-temporal attention.The lane mark area is pre-aligned by projecting the ground of the previous frame onto the current frame through feature point matching.A spatio-temporal attention module is designed to adaptively fuse multi-frame information,and the specific structure of the module and the way of attention generation are explored.The experimental results show that the robustness of lane mark detection can be effectively improved by this alignment fusion method.3.To deal with poor lane mark detection in low-light conditions such as shadows and nighttime,this thesis proposes a lane mark detection method based on unsupervised anti-forgetting domain adaptation.The unlabeled low-light images are adopted as the target domain for adversarial learning to generate robust domain-invariant features.Meanwhile,the reconstruction loss of constrained domain-specific feature learning is introduced to reduce the source domain feature forgetting risk.The experimental results show that the method effectively improves the detection performance of low-light scenes without degrading the source domain performance,and the overall performance is competitive with other existing methods.A series of lane mark detection methods are proposed in this paper,from using a single image as input,to multi-frame information fusion,and then to further expansion of the working domain.The effectiveness of the methods is fully verified by conducting experiments on multiple datasets,and the improvement of lane mark detection performance in complex scenarios is achieved.