Robust Lane Detection by Cascaded Deep Learning Network

Detecting lane markings in road scenes poses a significant challenge due to their intricate nature, which is susceptible to unfavorable conditions. While lane markings have strong shape priors, their visibility is easily compromised by varying lighting conditions, adverse weather, occlusions by other vehicles or pedestrians, road plane changes, and fading of colors over time. The detection process is further complicated by the presence of several lane shapes and natural variations, necessitating large amounts of high-quality and diverse data to train a robust lane detection model capable of handling various real-world scenarios.In this paper, we present a novel self-supervised learning method termed Contrastive Learning for Lane Detection via Cross-Similarity (CLLD) to enhance the resilience and effectiveness of lane detection models in real-world scenarios, particularly when the visibility of lane markings are compromised. CLLD introduces a novel contrastive learning (CL) method that assesses the similarity of local features within the global context of the input image. It uses the surrounding information to predict lane markings. This is achieved by integrating local feature contrastive learning with our newly proposed operation, dubbed cross-similarity.The local feature CL concentrates on extracting features from small patches, a necessity for accurately localizing lane segments. Meanwhile, cross-similarity captures global features, enabling the detection of obscured lane segments based on their surroundings. We enhance cross-similarity by randomly masking portions of input images in the process of augmentation. Extensive experiments on TuSimple and CuLane benchmark datasets demonstrate that CLLD consistently outperforms state-of-the-art contrastive learning methods, particularly in visibility-impairing conditions like shadows, while it also delivers comparable results under normal conditions. When compared to supervised learning, CLLD still excels in challenging scenarios such as shadows and crowded scenes, which are common in real-world driving.

While deep learning-based approaches have notably advanced the accuracy of lane detection in recent years, they remain limited in multi-scale feature modeling and the capture of fine-grained details. Methods such as CLRerNet model, which integrate refinement strategies and intersection over union (IoU) based loss optimization, have delivered strong performance; however, they still exhibit shortcomings in suppressing false positives and negatives, effectively fusing features, and optimizing overall network architecture. To address these challenges and improve robustness in complex driving environments, we propose an enhanced framework built upon CLRerNet, called ECLRerNet. The model introduces the MaxECA module into the segmentation branch—an enhanced efficient channel attention mechanism to enable adaptive multi-scale feature fusion, thereby improving the extraction of subtle lane structures. Furthermore, a large-kernel grouped attention gate (LGAG) module is designed to alleviate information imbalance during partial multi-scale fusion within the feature pyramid network. By dynamically regulating feature flow through an attention-guided mechanism, LGAG module reduces feature conflicts and enhances robustness. In addition, an improved IoU, termed ELaneIoU, is integrated into the loss function to suppress both false positives and false negatives by reducing the influence of erroneous lane points on loss computation. Extensive experimental results on the CurveLanes and CULane datasets demonstrate that ECLRerNet model achieves substantial performance improvements, reaching an <italic>F</italic>₁-score of 86.77% on CurveLanes and 81.58% on CULane—outperforming all existing open-source lane detection models on the CULane benchmark. These results verify the potential of ECLRerNet for reliable lane detection in real-world driving scenarios.

Accurate and robust lane detection is critical for autonomous driving systems, especially under complex road conditions and varying visual environments. This paper proposes a novel lane detection framework that integrates a Multi-Scale Residual Lane Graph Module (MSR-LGM) into the Polar R-CNN architecture. MSR-LGM addresses key limitations of conventional graph-based reasoning in lane detection by incorporating three core components: Multi-Scale Node Embedding, Enhanced Edge Encoding, and Feature Preservation via Residual Enhancement. These components collectively enhance lane feature representation, edge connectivity modeling, and learning stability. Experiments conducted on benchmark datasets—CULane—demonstrate that our method achieves state-of-the-art performance, outperforming prior works in accuracy and efficiency while maintaining a lightweight model size. The proposed method shows superior generalization across challenging scenarios such as occlusion, night, shadow, and lane curvature, making it well-suited for real-world autonomous driving applications.

Deep-learning-based lane detection degrades markedly in complex real-world conditions such as occlusion, low illumination, and missing markings. Although Ultra Fast Structure-aware Deep Lane Detection (UFLD) attains real-time inference through a lightweight design, its residual backbone remains deficient in multi-scale representation and deformable-geometry modelling, resulting in sub-optimal edge feature capture and, consequently, constrained detection accuracy and robustness. To enhance UFLD’s accuracy and robustness in complex scenarios, we propose fusing Multi-scale Convolutional Attention (EMCAM) and Deformable Convolutional Networks (DCN) in the feature extraction stage for improved feature utilization and shape modeling. Additionally, the Densely Connected Atrous Spatial Pyramid Pooling(DenseASPP) module is incorporated into the auxiliary segmentation for effective multi-scale contextual information capture. Ablation experiments validate these enhancements. Our improved UFLD achieves 96.03% F1 on TuSimple and 70.3% on CULane under real-time conditions. The proposed model enhances the original UFLD by remedying its feature-representation limitations, thereby achieving statistically significant improvements on established benchmark datasets while preserving its real-time inference capability.

&lt;div class="section abstract"&gt;&lt;div class="htmlview paragraph"&gt;Lane detection plays a critical role in autonomous vehicles for safe and reliable navigation. Lane detection is traditionally accomplished using a camera sensor and computer vision processing. The downside of this traditional technique is that it can be computationally intensive when high quality images at a fast frame rate are used and has reliability issues from occlusion such as, glare, shadows, active road construction, and more. This study addresses these issues by exploring alternative methods for lane detection in specific scenarios caused from road construction-induced lane shift and sun glare. Specifically, a U-Net, a convolutional network used for image segmentation, camera-based lane detection method is compared with a radar-based approach using a new type of sensor previously unused in the autonomous vehicle space: radar retro-reflectors. This evaluation is performed using ground truth data, obtained by measuring the lane positions and transforming them into pixel coordinates. The performance of each method is assessed using the statistical &lt;i&gt;R&lt;/i&gt;&lt;sup&gt;2&lt;/sup&gt; score, indicating the correlation between the detected lane lines and the ground truth. The results show that the U-Net camera-based method exhibits limitations in accurately detecting and aligning the lane lines, particularly in challenging scenarios. However, the radar-based lane detection method demonstrates a strong correlation with the ground truth which implies that the use of this sensor may improve current reliability issues from conventional camera lane detection approach. Furthermore, the study highlights the limitations of the U-Net model for camera lane detection, especially in scenarios with sun glare. This study shows that infrastructure-based radar retro-reflectors can improve autonomous vehicle lane detection reliability. The integration of different sensor modalities and the development of advanced computer vision algorithms are crucial for improving the accuracy, reliability, and energy efficiency of lane detection systems. Addressing these challenges contributes to the advancement of autonomous vehicles and the realization of safer and more efficient transportation systems.&lt;/div&gt;&lt;/div&gt;

Existing smart vehicles heavily depend on success of precise positioning, optical radar, visual detection and recognition to determine their road conditions and perfect routes. The visual-based approach is the simplest and most effective enabling technology to reach the goal. In performing such an overtaking maneuver, this vision-based driving assistant system must be able to precisely recognize the lane marks first. The traditional approach needs a classifier with hand-crafted features and an adjusted threshold to achieve a robust lane detection in various environment conditions. In this paper, we adopt the deep learning approach to achieve the robust lane detection. The proposed lane detection network inspired by LaneNet model, which uses semantic segmentation concepts utilizes multiple level features of the encoder and designs enhanced binary segmentation and reduced pixel embedding branches. By reduction of computation in decoders, the proposed network effectively utilizes multilevel features to precisely predict the high quality lane maps. The experiments on Tusimple and CuLane datasets verify that the proposed lane detection network achieves better accuracy performance than LaneNet and faster computation than the existed methods for real-time applications.

Lane marking detection is challenging under adverse weather, especially in rain. Rain affects the visibility by changing the amount of light reflected from the road and splashing of rain periodically blocks vision through the car windshield. Colored (yellow, blue etc.) and eroded lane markings have lower contrast between lane and road surface. Rain affects the visibility of these colored and eroded lanes severely. In this paper, a robust lane detection and tracking method is proposed to detect lane under heavy rainfall condition. Here, an automatic adjustable threshold (AAT) is proposed, which improves the performance of canny edge detector to detect colored, eroded and blurred lane markings. Probabilistic Hough Transform (PHT) is used to detect candidate lane lines. A lane verification technique slope and angle based geometric constraints (SAGC) is used to detect and verify lane markings. Finally, a lane tracking technique by angle range validation (ARV) is introduced, which can keep track of the lane position when either left or right or both lane markings are partially or fully invisible due to erosion or occlusion or when lanes are not even painted on the road. Experimental results show that the average detection rate is 94.12%, and the average detection time is 30.24ms per frame which outperformed state-of-the-art method.

Lane detection methods based on convolutional neural network have achieved excellent performance in recent years. Most of them treat lane detection as a semantic segmentation task which judges whether each pixel belongs to a lane. To make full use of the characteristics of lane shape, some researchers proposed to predict the whole lane. In this paper, we propose Lane Detection Combining Details and Integrity (LDCDI) which can explicitly leverage the advantages of both the segmentation-based methods and the regression-based methods. Specifically, we exploit an extra branch with regression-based methods as the auxiliary module after the main module. It not only maintains the advantages of the segmentation-based methods in lane detail segmentation, but also enables the model to have a sufficient understanding of the lane shape. Besides, the auxiliary module only takes part in the training, and there is no extra cost in the prediction. To further improve the quality of lane detection, we introduce a novel direction-sensitive block (DSB) based on ERFNet as the main module, which is more sensitive to the direction information of the image, so as to obtain better performance. Extensive experiments on the CULane dataset can demonstrate that our method outperforms other methods and achieves the state-of-the-art.

With the development of network and related technology has become more mature, now we can see the network applications in everywhere of the community, such as ATMs, mobile phones, home appliances and so on. With the increase of network applications as well as technology becomes more mature, we can foresee the future of the Internet is limitless. Network can be divided into wired network and wireless network, wireless network compared to a wired network has many advantages, such as easy to install, cover a wide range, amplified easy and so on. Wireless network thus ideal for use in difficult line installation place, such as vehicle, transportation node, institution and so on. Car is the most common means of transport, along with the progress of the automotive industry as well as increase of number of vehicles, how to enhance the safety of the vehicle is a very serious issue. Regarding to the safety of vehicle, in addition to considering the structural safety of the car, provide reliable auxiliary mechanism to driver is also a very important part. Millions of people injured or killed in a car accident, it is obvious that the issue of vehicle safety can not to be sniffed at. Because of the issue, the world-side countries are actively looking for solution to enhance the safety of vehicle. For example, the U.S. Department of Transportation’s National Highway Traffic Safety Administration (NHTSA) issued a final rule requiring all vehicles under 10,000 pounds, including buses and trucks, manufactured on or after May 1, 2018, to come equipped with rear visibility technology. Therefore, the rear visibility technology is a topic worthy of attention. In this paper, in order to decrease the technical threshold of establish the rear visibility system and avoid the hard installation to cause that can’t obtain the safety of vehicle. Therefore, we established a rear visibility system based on wireless transmission. By means of wireless transmission to increase the convenient of wiring and proposed a related power save mechanism to enhance the life time of wireless device in limited power. In addition, in order to avoid the system real time reaction capability decrease with power save mechanism, we proposed another auxiliary mechanism to reduce the effect that cause by power save mechanism. In addition, in order to avoid immediate saving mechanism to reduce system reaction capability, and the other proposed mechanisms to reduce the impact of an auxiliary power mechanism of response capabilities. The results show through simulations, through the mechanism proposed in this paper, the use of different scenario can effectively increase the life time of power and maintain a considerable degree of responsiveness.

Enhanced Cross Layer Refinement Network for robust lane detection across diverse lighting and road conditions

Lane detection plays a vital role in making the idea of the autonomous car a reality. Traditional lane detection methods need extensive hand-crafted features and post-processing techniques, which make the models specific feature-oriented, and susceptible to instability for the variations on road scenes. In recent years, Deep Learning (DL) models, especially Convolutional Neural Network (CNN) models have been proposed and utilized to perform pixel-level lane segmentation. However, most of the methods focus on achieving high accuracy while considering structured roads and good weather conditions and do not put emphasis on testing their models on defected roads, especially ones with blurry lane lines, no lane lines, and cracked pavements, which are predominant in the real world. Moreover, many of these CNN-based models have complex structures and require high-end systems to operate, which makes them quite unsuitable for being implemented in embedded devices. Considering these shortcomings, in this paper, we have introduced a novel CNN model named LLDNet based on an encoder–decoder architecture that is lightweight and has been tested in adverse weather as well as road conditions. A channel attention and spatial attention module are integrated into the designed architecture to refine the feature maps for achieving outstanding results with a lower number of parameters. We have used a hybrid dataset to train our model, which was created by combining two separate datasets, and have compared the model with a few state-of-the-art encoder–decoder architectures. Numerical results on the utilized dataset show that our model surpasses the compared methods in terms of dice coefficient, IoU, and the size of the models. Moreover, we carried out extensive experiments on the videos of different roads in Bangladesh. The visualization results exhibit that our model can detect the lanes accurately in both structured and defected roads and adverse weather conditions. Experimental results elicit that our designed method is capable of detecting lanes accurately and is ready for practical implementation.

Lane detection based on semantic segmentation can achieve high accuracy, but, in recent years, it does not have a mobile-friendly cost, which is caused by the complex iteration and costly convolutions in convolutional neural networks (CNNs) and state-of-the-art (SOTA) models based on CNNs, such as spatial CNNs (SCNNs). Although the SCNN has shown its capacity to capture the spatial relationships of pixels across rows and columns of an image, the computational cost and memory requirement needed cannot be afforded with mobile lane detection. Inspired by the channel attention and self-attention machine, we propose an integrated coordinate attention (ICA) module to capture the spatial-wise relationships of pixels. Furthermore, due to the lack of enhancement in the channel dimension, we created an efficient network with a channel-enhanced coordinate attention block named CCA, composed of ICA and other channel attention modules, for all-dimension feature enhancement. As a result, by replacing many repeated or iterative convolutions with the attention mechanism, CCA reduces the computational complexity. Thus, our method achieves a balance of accuracy and speed and has better performance on two lane datasets—TuSimple and ILane. At less than a few tenths of the computational cost, our CCA achieves superior accuracy compared to the SCNN. These results show that the low cost and great performance of our design enable the use of the lane detection task in autopilot scenarios.

Few-shot learning in deep networks through global prototyping

As is known to all, there are thousands of people died in traffic accidents in the world every year and 80% of the traffic accidents are due to the drivers' inattention. In order to reduce the occurrence of car accidents, the design of Cognitive Vehicle Safety (CVS) system has attracted wide attention. As an important functional system of the CVS system, Lane Departure Warning system's function is to warn the driver when the vehicle changes the lane without the driver's intention. In this paper, we present a method of lane detection which based on Lane geometric model and image connected region process. The model of the lane is produced by the projection of camera, which makes the parallel lane lines in real world become intersect. The lines on the same side of lane have the same geometric feature that can be used to build lane line geometric model for detection. The lane line can be classified as different lane line models by the geometric feature of it. In this paper, due to the different features of lane line in different images, the operation and the classification of modeling performs at the same time. But before that, we go through some process to pre-process the image, such as the image connected region process, morphology tophat filter, etc. The combination of these modules can overcome the universal lane detection problems, such as the road in different light conditions. Experimental results on real road will be presented to prove the effectiveness of the proposed lane detection algorithm.

According to the National Highway Traffic Safety Administration, one in ten fatal crashes and two in ten injury crashes were reported as distracted driver accidents in the United State during 2014. In an attempt to mitigate these alarming statistics, this paper explores using a dashboard camera along with computer vision and machine learning to automatically detect distracted drivers. We consider a dataset that incorporates drivers engaging in seven different distracting behaviors using left and/or right hands. Traditional handcrafted features paired with a Support Vector Machine classifier are contrasted with deep Convolutional Neural Networks. The traditional features include a blend of Histogram of Oriented Gradients and Scale-Invariant Feature Transform descriptors used to create Bags of Words. The deep convolutional methods use transfer learning on AlexNet, VGG-16, and ResNet-152. The results yield 85% accuracy with ResNet and 82.5% accuracy with VGG-16, which outperformed AlexNet by almost 10%. Replacing the fully connected layers by a Support Vector Machine classifier did not improve the classification accuracy. The traditional features yielded much lower accuracy than the deep convolutional networks.