Fang Gao

Tiny object detection is essential for railway intrusion monitoring, where early-stage targets often appear as only a few pixels at long distances. Low-light and adverse weather conditions further weaken their visual cues, causing feature loss, localization instability, and missed detections in conventional detectors. Here, we propose SMix-YOLOv8n, a unified framework for weather-aware and scale-adaptive tiny object detection in railway intrusion monitoring. First, a weather-aware feature enhancement is applied at the input stage to compensate for illumination and weather-induced degradation, preserving shallow cues critical for tiny object perception. Building upon this stabilized representation, a high-resolution detection pathway (Tiny-Head) retains fine-grained spatial details, while a scale-adaptive backbone enhancement (C2f_MixStructure) prevents tiny object features from being overwhelmed during deep feature aggregation. With these enriched representations, a newly designed shape-aware and scale-adaptive localization strategy (Mix_IoU) enables more stable bounding-box regression for visually degraded tiny targets. Experiments on a RailSem19-derived intrusion detection subset and VisDrone2019 demonstrate that SMix-YOLOv8n improves tiny object detection accuracy under adverse visual conditions. Compared with the YOLOv8n baseline, SMix-YOLOv8n improves mAP@0.5 from 65.6% to 69.5%, while maintaining an inference speed of 54 FPS. These results indicate that SMix-YOLOv8n achieves a practical balance between detection accuracy and inference efficiency, making it suitable for safety-critical railway intrusion monitoring under degraded visibility.

Accurate estimation of inter-target distance and azimuth is essential for safety-critical Unmanned Aerial Vehicle (UAV) applications such as rescue, surveillance, and navigation. Existing visual measurement systems typically estimate distance as an intermediate perception output and rarely treat azimuth as an explicitly defined measurand within a unified analytical framework, thereby limiting interpretability and uncertainty traceability. This paper presents UAVMD, a measurement-centric monocular projection geometry framework that enables traceable and uncertainty-aware joint distance-azimuth estimation. The method establishes closed-form analytical mappings from pixel coordinates, UAV altitude and pitch angle, and calibrated camera parameters to metric outputs, forming a transparent and traceable measurement chain. Terrestrial experiments conducted at altitudes of 639-1799 m and inter-target distances of 36-82 m achieved a distance MAE of 2.92 m and RMSE of 3.45 m. Maritime infrared experiments conducted at an altitude of 600 m with target separations of 22-45 m achieved a distance MAE of 0.481 m and azimuth MAE of 0.286°. The computational latency is 0.24 ms per measurement. Sensitivity and uncertainty analyses confirm predictable error propagation under platform disturbances. The proposed framework provides an accurate, lightweight, and real-time deployable monocular solution for UAV-based geometric measurement tasks.