Purpose The purpose of this paper is to design a fully supervised point cloud voxel network method for foreign object intrusion detection, which integrates CSF filtering and can effectively address the issue where radar data interference from railway track reflections affects the detection of small objects. Design/methodology/approach A supervised learning-based method for foreign object intrusion detection is proposed, which integrates a voxel-based point cloud network with the Cloth Simulation Filtering (CSF) algorithm. Findings The CSF algorithm is applied to filter out ground points from the railway scene, effectively mitigating the interference caused by strong radar reflections from the track surface, which often hinder the detection of small objects. Considering railway-specific scene characteristics, an improved Voxel R-CNN is introduced: the method first performs voxelization and voxel-wise feature encoding, then uses a three-dimensional (3D) convolutional backbone for multiscale feature fusion. The resulting features are compressed via Bird’s Eye View projection, and a voxel query mechanism is used to generate 3D bounding boxes for regions of interest. Research limitations/implications Despite the promising performance of the proposed method, several limitations should be acknowledged. First, the detection accuracy of small objects (e.g. stones) remains susceptible to point cloud sparsity and LiDAR resolution, especially under long-range sensing conditions. Second, the current model is trained and validated exclusively on data collected from a controlled railway test environment, and its generalization capability to diverse track structures, sensor configurations and environmental conditions requires further verification. Notably, this study does not include a direct experimental comparison with video-based detection methods under adverse weather conditions. Although prior studies have demonstrated the superior robustness of LiDAR sensing in low-illumination and visually challenging scenarios, a rigorous cross-modality comparison necessitates synchronized multisensor data collected under identical weather conditions – an issue that will be addressed in future work. In future work, the authors plan to incorporate multisensor fusion and domain adaptation strategies to enhance robustness across diverse railway scenarios. In addition, lightweight model optimization will be explored to facilitate real-time deployment on embedded railway monitoring platforms. Originality/value Experimental results show the method achieves average detection accuracies of 80% for pedestrians and stones, and 40% for boxes, demonstrating its effectiveness for on-site deployment.
Lin et al. (Mon,) studied this question.