Millimeter-Wave Radar Target Detection Using Pointnet++ Combined with Multi-Scale Feature Fusion

Millimeter-wave radar point clouds are sparse and noisy, posing challenges for accurate target detection. This paper proposes a detection method combining PointNet++ with an attention-guided multi-scale feature fusion strategy. Using farthest point sampling and spherical queries, multi-scale local features are extracted and encoded. An attention-based fusion module enhances feature representation by adaptively weighting channels across scales. Preprocessing with density filtering improves input quality, while classification and regression branches output object categories and spatial parameters. Experimental results show improved detection accuracy, achieving 76.2 % for pedestrians and 86.5 % for cars. Under high-intensity noise, the method maintains robust accuracy of 0.74, demonstrating enhanced resilience and generalization. This approach addresses data sparsity and noise issues and supports the deployment of low-cost sensors in complex environments.