A Fusion of Graph- and Grid-Based Hybrid Model of Object Detection and Semantic Segmentation for 4-D Millimeter-Wave Radar

The next-generation 4-D millimeter-wave radar can provide rich information and dense point cloud and perceive the environment under all-weather and all-operating conditions, making it very suitable for autonomous driving systems. However, the existing object detection and semantic segmentation models based on 4-D millimeter-wave radar are usually directly transplanted from light laser detection and ranging (LiDAR), which cannot effectively adapt to the millimeter-wave radar point cloud features, resulting in poor detection performance. Concerning this, a hybrid model for 4-D millimeter-wave radar object detection and semantic segmentation is developed here via fusing graph and grid. A graph neural network (GNN) module named radar adaptive multichannel GNN (RAMGNN) is first designed, which leverages topology graphs and feature maps to propagate and update point cloud features. The node embeddings outputted from RAMGNN can be directly used for semantic segmentation and serve as a point cloud feature encoder for subsequent object detection. In what follows, the point cloud is projected onto a 2-D bird's-eye view (BEV) grid, and its multiscale features can be extracted exploiting a backbone network with channel attention mechanism. Finally, multiscale features are fused to achieve effective object detection and semantic segmentation. Experimental results conducted on the publicly available dataset view-of-delft (VoD) demonstrate that the proposed model outperforms state-of-the-art algorithms in terms of both object detection performance and semantic segmentation quality.