Tail point filtering for industrial LiDAR using angular projection and dual spatial neighborhood thresholds

Non-repetitive solid-state LiDARs such as Livox Mid-360 are widely deployed on Automated Guided Vehicles (AGVs) but suffer from severe “tail point” artifacts at object boundaries, degrading pose estimation and safetycritical perception. Existing geometric filters are either too slow for embedded CPUs or fail to distinguish spurious tails from true structural edges. We propose AP-DST (Angular Projection with Dual Spatial Thresholding), a lightweight tail point filter that projects the cloud onto a 2D angular grid and applies dual-stage spatial thresholding followed by point-wise depth rejection. Height discontinuities first locate candidate regions; a continuity measure then separates structural edges from isolated tails; finally, a depth threshold removes trailing points along the LiDAR ray. Experiments on a Mid-360–equipped forklift in real warehouse scenes show that AP-DST achieves >95% tail removal with >91% structure preservation, while running 3–10× faster than common baselines on CPU-only hardware.