Thursday 23 January 2025
As autonomous vehicles hit the roads, one of the biggest challenges facing developers is how to accurately identify and classify objects in real-time using radar data. Radar sensors have the potential to provide a wealth of information about the environment around a vehicle, including speed, distance, and direction of movement. However, processing this data into meaningful labels for use by autonomous systems has proven difficult.
A new approach published in a recent journal article tackles this problem head-on by proposing an automatic labelling process that uses complementary information from cameras and LiDAR sensors to generate point-by-point multi-class labels. These labels can then be used as ground truth with corresponding 4D radar data to train a semantic segmentation network, which associates class labels to each spatial voxel.
The authors of the article begin by using a pre-trained object detection model to generate rough labels on LiDAR point clouds (PCs). They then use RGB images from cameras to calibrate key labels in the central view and adjust label consistency through a clustering algorithm. The resulting labels are transformed into a cube with dimensions 500x240x34, representing range, azimuth, and elevation.
The proposed semantic segmentation network consists of two individual branches: one for generating 3D occupancy features and another for classifying targets. These feature maps are then combined to form a 3D semantic information latent space, which is used as input for the final classification step.
In tests on the publicly shared RaDelft dataset, the proposed approach achieved over 65% of the LiDAR detection performance, improving 13.2% in vehicle detection probability and reducing 0.54 meters in Chamfer distance compared to variants inspired from the literature.
The authors also compared their approach with two other methods: one that uses a sparse representation of radar data and another that adapts a deep learning model for object detection into a semantic segmentation network. While these approaches showed some promise, they fell short of the proposed method’s performance.
This work has significant implications for the development of autonomous vehicles, as it demonstrates a reliable and efficient way to process radar data into meaningful labels. The authors’ approach could be used in a variety of applications, from object detection and tracking to scene understanding and prediction.
The use of automatic labelling techniques also opens up new possibilities for training machine learning models on large datasets without the need for manual annotation. This could accelerate the development of autonomous systems and enable them to operate more effectively in complex environments.
Cite this article: “Automated Radar Data Labelling for Autonomous Vehicles”, The Science Archive, 2025.
Radar Data, Autonomous Vehicles, Object Detection, Semantic Segmentation, Lidar Sensors, Camera Data, Machine Learning Models, Automatic Labelling, Ground Truth Labels, 4D Radar Data
