Get Started with Lidar Lane Detection Using Deep Learning

Lane Detection refers to estimating the accurate location and the curvature of road lanes. It is a crucial step in navigation and path planning.

You can train a deep learning network to detect road lanes using the data captured by sensors like a camera and lidar. A lane detection network must be robust with respect to occlusions, as well as adverse lighting and weather conditions.

Lidar lane detection (LLD) networks have several advantages over networks that use other sensors, as they are unaffected by environmental conditions and provide accurate results even over very long distances. Automated Driving Toolbox™ provides a lidar lane detection network trained on the K-Lane data set which is the largest lidar lane data set containing urban roads and highways, in the world.

LLD-GFC Network

A lidar lane detection network utilizing global feature correlator (LLDN-GFC) network detects road lanes from a lidar point cloud data by exploiting the spatial characteristics of lane lines in the point cloud.

Convolutional neural network (CNN) based LLD networks use front-view images of the lane lines, where the lane thickness vanishes with increased distance from the ego vehicle. Hence, CNN-based networks cannot effectively use the spatial characteristics of the lane lines. On the other hand, the LLDN-GFC network uses bird's-eye-view (BEV) images, where the lane lines maintain a constant thickness throughout the stretch of the road. These BEV images enable effective feature extraction, and enable the network to generate robust detections.

The LLD-GFC networks consists of these parts.

1. BEV Encoder — Projects 3-D point cloud data onto a 2-D image, and further processes it to generate a 2-D BEV feature map.

2. Global Feature Correlator — Computes correlations between distance grids within the feature map using patch-wise self-attention networks.

3. Detection Head — Detects lane classes of various shapes using multiclass segmentation. The detection head consists of two segmentation heads, each containing a shared multilayer perceptron with nonlinear activation.

Create Lidar Lane Detection Network

To programmatically create a lidar lane detection network, use the lidarLaneDetector object.

Train and Detect Lanes in Lidar Data

Use the trainLidarLaneDetector function to train a lidar lane detector network. To perform lane detection with a trained lidar lane detector network, use the detect function.

Evaluate Lane Detection Results

Use the evaluate function to evaluate the detection results generated by the lidarLaneDetector network object.

References

[1] Paek, Dong-Hee, Seung-Hyun Kong, and Kevin Tirta Wijaya. “K-Lane: Lidar Lane Dataset and Benchmark for Urban Roads and Highways.” In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 4449–58. New Orleans, LA, USA: IEEE, 2022. https://doi.org/10.1109/CVPRW56347.2022.00491.