RailMASDet: a 3D object detection method for railway LiDAR point cloud foreign object detection

Abstract To address the issues of easy loss and inaccurate localization of distant small targets in railway LiDAR point cloud foreign object detection, this paper proposes a 3D object detection method named RailMASDet. The method constructs a cross-scale semantic spatial feature reassembly strategy. By employing transposed restoration of deep features and aligned compression of shallow features, it resolves the problem of spatial detail loss of distant targets in deep BEV feature maps. Furthermore, a railway-oriented channel modulation strategy based on existing channel attention mechanisms is introduced. By jointly exploiting global average and maximum statistics, the strategy adaptively enhances object-sensitive responses and suppresses structured-background interference in railway BEV features. Meanwhile, a foreground probability-driven non-Euclidean distance sampling paradigm is established, embedding point-level foreground priors into the distance metric function to shift keypoint sampling from uniform coverage to semantically guided targeted focusing. On the OSDaR23 dataset, RailMASDet achieves AP 0.7 of 80.59% for pedestrians and 70.05% for car, and an mAP 0.5 of 76.56%, outperforming the baseline model by 4.95, 3.78, and 3.21 percentage points, respectively. The average center localization error of detection boxes is 0.86 m, which is 57.5% lower than that of the baseline. The inference latency is 276.5 ms per frame, demonstrating reasonable computational efficiency. Ablation studies and visualization analyses validate the individual contributions of the three modules and their synergistic enhancement effects, confirming that the proposed method can effectively solve the problems of small target loss and insufficient localization accuracy in railway foreign object detection, thereby demonstrating its potential for long-range foreign object detection in sparse railway LiDAR scenes.