OBSTACLE DETECTION ON RAILWAY TRACKS FROM UAV IMAGERY: A LEAKAGE-AWARE EVALUATION

Abstract

On industrial sidings and station tracks, foreign objects and people create a collision risk during shunting, and automatic analysis of images taken by an unmanned aerial vehicle (UAV) could speed up track inspection. Such detectors are evaluated on the split published with an open dataset, but whether the test images are independent of the training images is rarely checked. This study measures data leakage in the published split of the open UAV-RSOD dataset (2,002 images, six obstacle classes) and its effect on the estimated detector quality. Near-duplicates were searched with SIFT keypoints and RANSAC geometric verification; a blocked split that reduces near-duplicates was built from the near-duplicate graph with a buffer zone; a YOLOv8n detector was fine-tuned with identical settings on a random and a blocked split of equal size; COCO-pretrained YOLOX detectors were tested on the person class without fine-tuning. Of the published test images, 390 of 391 (99.7 %) have a RANSAC-verified near-duplicate (augmented copy or adjacent frame) in the training set. Compared with a random split of equal size, the blocked split lowered the share of test images with a same-class near-duplicate from 99.8 % to 64.8 % but could not remove it, because all images come from one track section. YOLOv8n reached mAP50 = 0.806 on the random split and 0.719 on the blocked split (one run); the difference varied across classes. The best zero-shot AP50 for people seen from above was 0.455. All images come from one track section with staged obstacles, and false-alarm rates could not be estimated because the dataset has no obstacle-free images. The contribution is a quantified leakage estimate for UAV track-inspection data and a procedure to reduce it; in practice, detectors should be accepted on track sections not used for training.