Radar Information Transformer: A Sparse Single Frame-Based Classification Network for mmWave Radar

3D object detection is critical for autonomous driving systems, particularly in adverse weather conditions where 4D radar offering both spatial and velocity data outperforms traditional sensors. Despite its advantages, 4D radar struggles with noise and ambiguity in measurement. Existing detection methods fail to identify target objects due to their reliance exploring Local Features within sparse point clouds. To address these challenges, we propose the Radar Information Transformer Network (RITnet), which employs a feature-wise attention mechanism to capture the overarching characteristics of point clouds from 4D radar data, rather than relying on PointNet. We train our network on a proprietary radar classification dataset and the TJ4DRadSet dataset. Experimental results indicate that the overall classification accuracy reaches 97.5%, and exhibits good performance in terms of F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> scores and Hamming-loss compared to popular deep learning frameworks.