FIXED-THRESHOLD RGB–POSE FALL DETECTION ACROSS PARTICIPANTS UNDER CONTROLLED IMAGE DEGRADATION
Keywords:
Fall detection, RGB video, Human pose, Image degradation, Cross-participant evaluationAbstract
This study evaluates how fixed-threshold RGB–pose fall detection transfers across participants and responds to controlled image degradation. On UP-Fall, participants are separated into fitting, confirmation and final evaluation groups. A frozen image encoder and a geometric temporal pose classifier are combined by equal probability averaging at threshold 0.5. On final clean trials, the predefined primary seed achieves 94.42% window recall, 1.54% normal-window false-positive rate and 39/39 event hits. On matched trials, recall falls from 90.95% to 87.16% under darkening and 73.49% under strong blur. Across three seeds, mean blur recall is 44.87%, while valid pose geometry covers 19.91% of blurred sampled frames. The evaluation quantifies fixed-threshold tradeoffs and shows why event hits should be reported alongside window recall, seed sensitivity and pose coverage.References
[1] Qi L, Sun H. Multistage fall detection framework via 3D pose sequences and TCN integration. Scientific Reports, 2025, 15(1): 27832.
[2] Howard A, Sandler M, Chu G, et al. Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. 2019: 1314-1324.
[3] Jiao S, Li G, Zhang G, et al. Multimodal fall detection for solitary individuals based on audio-video decision fusion processing. Heliyon, 2024, 10(8): e29596.
[4] Rehouma H, Boukadoum M. Fall detection by deep learning-based bimodal movement and pose sensing with late fusion. Sensors, 2025, 25(19): 6035.
[5] Zobi M, Bolzani L, Tabii Y, et al. Robust 3D Skeletal Joint Fall Detection in Occluded and Rotated Views Using Data Augmentation and Inference–Time Aggregation. Sensors, 2025, 25(21): 6783.
[6] Pham Q N, Ninh D K. Lightweight fall detection on edge devices via knowledge distillation. The University of Danang – Journal of Science and Technology, 2025: 74-78.
[7] Zhao X, Jin P. Graph Topology-Guided Multi-Task Learning for Fall Detection. Sensors, 2026, 26(17): 5529.
[8] Hendrycks D, Dietterich T. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019.
[9] Martínez-Villaseñor L, Ponce H, Brieva J, et al. UP-fall detection dataset: A multimodal approach. Sensors, 2019, 19(9): 1988.
[10] OpenCV Zoo. Person detection and pose estimation MediaPipe model implementations, commit 47534e27c9851bb1128ccc0102f1145e27f23f98, 2026. https://github.com/opencv/opencv_zoo/tree/47534e27c9851bb1128ccc0102f1145e27f23f98/models.
[11] Lugaresi C, Tang J, Nash H, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019.
[12] Bazarevsky V, Grishchenko I, Raveendran K, et al. Blazepose: On-device real-time body pose tracking. arXiv preprint arXiv:2006.10204, 2020.