TransVoxelRadar: Transformer-Based Voxel Radar Perception for Autonomous Driving

Accurate and efficient 3D object detection is crucial for safe and reliable autonomous driving. While radar sensors offer robust distance and velocity measurements, their sparse and semantically limited data pose challenges for object detection. This paper introduces TransVoxelRadar, a novel approach that leverages transformer-based view transformation and voxel-based representations for 3D object detection using radar data. We propose a single-stage architecture that directly operates on raw radar point clouds, eliminating the need for hand-crafted feature engineering. Our method employs a transformer encoder to efficiently transform radar data into a voxel grid representation, capturing both local and global contextual information. We then utilize a voxel-based 3D convolutional neural network to detect objects directly in the voxel space. Extensive experiments on a synthetic dataset demonstrate the effectiveness of TransVoxelRadar, achieving competitive performance in terms of accuracy and inference speed compared to existing radar-based methods. Furthermore, our approach exhibits robustness to noise and occlusion, making it suitable for real-world autonomous driving scenarios.