Lane Boundary Detection for Intelligent Vehicles Using Deep Convolutional Neural Network Architecture

To address the limitation of 2D lane detection methods with monocular vision, which fail to capture the three-dimensional position of lane boundaries, this study proposes a convolutional neural network architecture for 3D lane detection. The deep residual network ResNet50 is employed as the feature extraction backbone, augmented with a coordinate attention mechanism to facilitate shallow feature extraction, multi-scale feature map generation, and extraction of small-scale high-order feature information. The BIFPN network is utilized for bidirectional feature fusion across different scales, significantly enhancing the accuracy of lane boundary detection. By constructing an inverse perspective transformation model (IPM), the conversion from front view to aerial view is realized. A dedicated 3D lane detection head is designed for lane boundary anchor lines, enabling efficient fusion and downsampling of multi-scale feature maps. By incorporating the bias between lane boundaries and anchor lines, the 3D position of lane boundaries is effectively detected. Validation experiments on the OpenLane dataset demonstrate that the proposed method not only detects the spatial locations of lane boundaries but also identifies attributes, such as color, solid or dashed, single or double lines, and left or right dashed configurations. Additionally, the method achieves an inference speed of 64.9 FPS on an RTX 4090 GPU, showcasing its computational efficiency.