PillarDANet: An Improved PillarNet with Dual-Attention Mechanism for Ship Detection

This paper proposes PillarDANet, an improved 3D point cloud ship detection method based on PillarNet, specifically targeting ship detection in high-head large ship locks. The method introduces a dual attention mechanism on PillarNet, including Position Attention Module (PAM) and Channel Attention Module (CAM), enhancing the perception of complex ship geometric structures. Additionally, multi-scale dilated convolution strategy is adopted in the neck network, significantly expanding the receptive field and improving the detection performance for large ship targets. Experimental results on the nuScenes dataset and a self-built ship dataset show that compared to the baseline PillarNet model, this method achieves 3.1% and 3.9% improvements in mAP metrics respectively, demonstrating effective detection performance.