Abstract:In the process of lane line detection algorithm implementation, traditional image processing techniques rely too much on manually designed features and parameters, which cannot meet the requirements of autonomous driving for detection accuracy in complex scenes such as lighting changes, severe occlusion, shadow interference, etc. Moreover, the curve modeling method cannot achieve ideal results for fitting lane lines in curved scenes. In order to detect and track lane lines quickly and accurately so as to improve the safety and reliability of autonomous driving, this paper proposes a fast lane detection algorithm that fusing multi-scale contour feature extraction and weight mean normalization, MREFE-UFLD-WMN. By using a combination of theoretical analysis, image processing and ablation experiments, the lane line detection model is constructed by considering attention mechanism, null convolution, multi-scale residual edge contour feature extraction and weighted mean normalization strategy to analyze the effects of increasing feature attention and field of view, enhanced object contour information extraction and weighted coefficient mean normalization on the lane line detection effect, following the principle of gradual detailing. The summation fusion of feature layers is carried out so that the model obtains rich edge information and assists the backbone network to extract image features so that the model has both detection accuracy and real-time performance. The experimental results show that the proposed detection algorithm can better balance detection speed and accuracy on the common dataset CULane (normal, crowded, night, occlusion, shadow, light, curve, loop, etc.) containing multiple scenes, and significantly improve the detection accuracy at the cost of increasing the number of parameters by a very small amount, compared with the benchmark model (ultra fast lane detection, UFLD), the accuracy (precision, P), recall (R) and F1 score (F1) are improved by 10.6%, 4.8% and 7.8%, respectively, which can adapt to the complex traffic environment and meet the accuracy and real-time of lane detection in complex scenes. It has high application value in scenarios such as vehicle–road vision collaboration, vehicle tracking and vehicle anti-collision warning.

Lane Detection Based on a Lightweight Convolutional Neural Network

DILane: Dynamic Instance-Aware Network for Lane Detection

Lane Detection Combining Details and Integrity: an Advanced Method for Lane Detection

Lane Detection in Complex Scenes Based on End-to-end Neural Network

A Lane Line Detection Algorithm Based on Convolutional Neural Network

Efficient Lane Detection Technique Based on Lightweight Attention Deep Neural Network

Fast lane detection by fusing multi-scale contour feature extraction and weight mean normalization

Shallow Detail and Semantic Segmentation Combined Bilateral Network Model for Lane Detection

DCLane for Real-time Understanding of Lane Markings.

Lightweight real-time lane detection algorithm based on ghost convolution and self batch normalization

Traffic Lane Detection Using Fully Convolutional Neural Network

A Fast and Robust Lane Detection via Online Re-Parameterization and Hybrid Attention

A Faster and Lightweight Lane Detection Method in Complex Scenarios

CondLaneNet: a Top-to-down Lane Detection Framework Based on Conditional Convolution

Efficient Multi-Lane Detection Based on Large-Kernel Convolution and Location

Combining Low-Light Scene Enhancement for Fast and Accurate Lane Detection

An Efficient Lane Detection Network with Channel-Enhanced Coordinate Attention

Asymmetric convolution multi lane detection and tracking network based on spatial semantic segmentation

Fast Lane Detection Method Based on Improved Multi-branch Residual Network

LaneNet: Real-Time Lane Detection Networks for Autonomous Driving

Fast-HBNet: Hybrid Branch Network for Fast Lane Detection