EATF: Evidence-aware temporal filtering for recovering low-confidence UAV detections
DOI:
https://doi.org/10.54939/1859-1043.j.mst.114.2026.159-166Keywords:
UAV detection; Temporal filtering; Low-confidence detections.Abstract
Small UAVs often produce weak and intermittent detections because of their small image size, background clutter, and rapid apparent motion. This paper presents Evidence-Aware Temporal Filtering (EATF), an interpretable offline post-processing method for a frozen detector. High-confidence detections are retained directly, whereas weaker detections are recovered only when their completed associated tracks satisfy fixed mean-confidence and cumulative-confidence thresholds. EATF does not retrain the detector, estimate motion states, interpolate boxes, or generate new detections. On DUT Anti-UAV Tracking, EATF increased frame-level recall from 0.4377 to 0.5319 and F1 from 0.5870 to 0.6416 relative to detector thresholding at 0.55. At matched recall, it improved precision from 0.7657 to 0.8084 and F1 from 0.6277 to 0.6416. Compared with default ByteTrack using post-tracking top-1 selection, EATF achieved higher recall and F1, with paired sequence-level bootstrap intervals supporting both gains. The current method remains retrospective, single-target, and vulnerable to persistent background false alarms under the evaluated DUT protocol.
References
[1]. A. Coluccia et al., “Drone vs. Bird Detection: Deep Learning Algorithms and Results from a Grand Challenge”, Sensors, Vol. 21, No. 8, p. 2824, (2021). DOI: 10.3390/s21082824.
[2]. A. Coluccia et al., “Drone-vs-Bird Detection Grand Challenge at ICASSP2023”, ICASSP 2023, IEEE, pp. 1–2, (2023).
[3]. F. N. M. Zamri et al., “Enhanced small drone detection using optimized YOLOv8 with attention mechanisms”, IEEE Access, Vol. 12, pp. 90629–90643, (2024).
[4]. N. V. Ngon et al., “Optimizing long-range UAV detection on YOLOv8: Breaking-point distance analysis and combining adaptive tiling with AdamW optimizer”, Journal of Military Science and Technology, Vol. 109, pp. 154–163, (2026). DOI: 10.54939/1859-1043.j.mst.109.2026.154-163.
[5]. F. Kuppers et al., “Multivariate Confidence Calibration for Object Detection”, Proceedings of the IEEE/CVF Conference on CVPRW, pp. 326–327, (2020). DOI: 10.1109/CVPRW50498.2020.00171.
[6]. U. Seidaliyeva et al., “Real-time and accurate drone detection in a video with a static background”, Sensors, Vol. 20, No. 14, p. 3856, (2020).
[7]. J. Zhu et al., “Bird and UAVs recognition detection and tracking based on improved YOLOv9-DeepSORT”, IEEE Access, Vol. 12, pp. 147942–147957, (2024). DOI: 10.1109/access.2024.3475629.
[8]. X. Cheng et al., “IRWT-YOLO: A Background Subtraction-Based Method for Anti-Drone Detection”, Drones, Vol. 9, No. 4, p. 297, (2025). DOI: 10.3390/drones9040297.
[9]. Y. Sun et al., “Enhancing UAV detection in surveillance camera videos through spatiotemporal information and optical flow”, Sensors, Vol. 23, No. 13, p. 6037, (2023).
[10]. L. K. Thanh et al., “Real-time long-term target tracking on an ARM platform with NPU acceleration and integration into UAV line-of-sight stabilization”, Journal of Military Science and Technology, Vol. 109, pp. 25–34, (2026). DOI: 10.54939/1859-1043.j.mst.109.2026.25-34.
[11]. Y. Zhang et al., “ByteTrack: Multi-object tracking by associating every detection box”, European Conference on Computer Vision, Springer, pp. 1–21, (2022).
[12]. J. Zhao et al., “Vision-Based Anti-UAV Detection and Tracking”, IEEE Transactions on Intelligent Transportation Systems, Vol. 23, No. 12, pp. 25323–25334, (2022). DOI: 10.1109/TITS.2022.3177627.
[13]. Ultralytics, “Ultralytics YOLO11”, Ultralytics Docs, Available: https://docs.ultralytics.com/mo-dels/yolo11
