Target Point Cloud Segmentation for Battery Swapping Robot Based on Multiscale Attention Aggregation

Rapid and accurate segmentation of 3D point clouds is critical for optimizing battery-swapping robots and ensuring precise assembly. To address the challenges of computational inefficiency and insufficient accuracy for small fasteners in large-scale scenes, this paper proposes MSA-RandLA-Net, a novel segmentation network based on multi-scale attention aggregation. Specifically, the method first employs a preprocessing scheme combining radius filtering and DBSCAN clustering to effectively eliminate background noise, thereby ensuring high-quality input data. Subsequently, a multi-scale attention aggregation module is embedded within the residual blocks of the network to adaptively capture fine-grained geometric details and enhance edge feature extraction. Experimental results on our real-world battery-swapping dataset demonstrate superior performance, achieving an mIoU of 94.2%, mAcc of 97.1%, and OAcc of 93.2%. Notably, the model maintains a lightweight design with only 1.02 M parameters and a 30 ms inference time, fully satisfying the stringent requirements for real-time battery swapping.