A multi-point focus transformer approach for large-scale ALS point cloud ground filtering

In recent years, Transformer networks have achieved a series of advancements in 3D point cloud semantic segmentation and shape classification. In this paper, we propose a multi-point focus transformer network for outdoor large-scale point cloud filtering. It integrates farthest point sampling and random sampling methods to extract both global and local multi-feature information from point clouds. To more accurately compute the self-attention and positional encoding of point clouds, this paper proposes a multi-point focus mechanism that uses a combination of farthest point sampling and random sampling to select multiple focal points from neighbourhoods at different scales for special focused, followed by attention computation and positional encoding for these focal points. Subsequently, an attention integration module is introduced to aggregate the self-attention and positional information from multiple focal points. Finally, the idea of inverse residual MLP was borrowed to obtain deeper level features of point clouds through extended channels. Extensive experiments were conducted on the latest OpenGF dataset for different terrain scenarios, resulting in commendable filtering accuracy. On the Test1 dataset, qualitative visual comparison and quantitative analysis were conducted with other state-of-the-art methods, and the overall accuracy (OA) could reach up to 98.12%, further verifying the effectiveness and competitiveness of the proposed multi-point focusing transformer network.