Researchers have designed FGLD-YOLO, a lightweight computer-vision model derived from YOLOv10 that detects fall-related human movement for assistive healthcare, rehabilitation, and older-adult monitoring. Practical deployment in these continuous monitoring settings requires achieving a favorable operational balance among detection accuracy, model complexity, and inference efficiency. To meet these real-world demands, the investigators engineered architectural modifications that reduce computational overhead while enhancing detection capabilities.
The foundational architecture of FGLD-YOLO modifies multiple core components of the standard YOLOv10 framework. The research team replaced the original backbone network with FasterNet to directly reduce computational overhead during feature extraction. To strengthen multi-scale feature fusion across diverse human postures, the developers incorporated a Generalized Feature Pyramid Network into the neck of the detector.
Following the backbone and neck redesign, the modified components were paired with a YOLOv8 detection head to complete a unified hybrid detection architecture.
To translate these structural improvements into a deployment-oriented system, the authors applied feature-based knowledge distillation. Under this framework, the optimized hybrid detector functioned as a teacher model, transferring its learned representations directly to a smaller YOLOv10n student network. This distillation process produced a compact detector capable of efficient inference without sacrificing significant analytical precision.
The experimental evaluation utilized 5,000 images compiled from two established open-source benchmarks: the UR Fall Detection Dataset, known as URFD, and the Fall Detection Dataset, known as FDD. The assembled dataset was partitioned into training, validation, and test sets according to a strict 7:2:1 ratio to systematically assess detection behavior under controlled experimental testing conditions.
In comparative testing against standard YOLOv10, the optimized full FGLD-YOLO model improved mAP50 by 4.2 percentage points, precision by 4.7 percentage points, and recall by 8.3 percentage points. The full detector also demonstrated competitive performance when evaluated against established alternate architectures, specifically YOLOv8 and YOLOv12.
The distilled student model designed for practical deployment achieved 91.9 percent mAP50, 87.9 percent precision, and 85.1 percent recall. These metrics represent gains of 3.1, 4.1, and 3.2 percentage points over baseline YOLOv10, confirming that the student retained key analytical capabilities from the larger hybrid teacher model during distillation.
When evaluated against the optimized full detector, the distilled student achieved substantial efficiency gains, reducing overall parameter count by 53.2 percent, lowering GFLOPs by 42.0 percent, and decreasing per-image processing latency by 56.0 percent. These findings confirm an improved trade-off between accuracy and computational efficiency on the tested datasets.
The authors conclude that these performance metrics support further investigation of FGLD-YOLO for deployment in resource-constrained monitoring and assistive-care applications. Nevertheless, the researchers note that extensive validation in real-world settings and across actual older-adult populations remains necessary before the framework can be deployed for clinical use.
Subject of Research: Vision-based human fall detection using lightweight deep learning models
Article Title: A lightweight and efficient model for fall-related human movement detection based on improved YOLOv10
Article References: Ma, C., Li, M., Mi, J., Yao, J., Su, P., & Zuo, Y. (2026). A lightweight and efficient model for fall-related human movement detection based on improved YOLOv10. Scientific Reports. https://doi.org/10.1038/s41598-026-79968-z
Image Credits: AI Generated
DOI: 10.1038/s41598-026-79968-z
Keywords: FGLD-YOLO, YOLOv10, fall detection, vision-based fall detection, human movement detection, FasterNet, Generalized Feature Pyramid Network, GFPN, YOLOv8, YOLOv12, knowledge distillation, UR Fall Detection Dataset, URFD, Fall Detection Dataset, FDD, assistive healthcare, older-adult monitoring, mAP50, inference efficiency, model compression
Cite Scienmag News
Blake Davidson. (November 21, 2026). Compact YOLOv10 Model Improves Human Fall Detection Efficiency in Initial Image Tests. Scienmag. https://scienmag.com/compact-yolov10-model-improves-human-fall-detection-efficiency-in-initial-image-tests/
Blake Davidson. "Compact YOLOv10 Model Improves Human Fall Detection Efficiency in Initial Image Tests." Scienmag, 21 November 2026, https://scienmag.com/compact-yolov10-model-improves-human-fall-detection-efficiency-in-initial-image-tests/. Accessed 21 November 2026.
Blake Davidson. "Compact YOLOv10 Model Improves Human Fall Detection Efficiency in Initial Image Tests." Scienmag. November 21, 2026. https://scienmag.com/compact-yolov10-model-improves-human-fall-detection-efficiency-in-initial-image-tests/








