Fig 1.
The proposed framework enhances YOLOv8n by integrating four key modules: the BiFormer attention for global feature awareness, C2f_rvb for multi-scale feature extraction, MultiSEAM attention in the detection head for occlusion robustness, and the SIoU loss for improved box regression. This design specifically targets challenges in complex fall detection scenarios, such as occlusion and low-light conditions.
Fig 2.
Structure of the BiFormer attention mechanism.
This module efficiently models long-range dependencies via a dual-layer routing process. It partitions features into regions, computes affinities, and applies top-k selection to route computation only to the most relevant regions, enabling dynamic sparse attention and reducing computational overhead.
Fig 3.
Workflow of the RepViT Block (RVB).
The RVB module replaces the standard bottleneck to enhance feature extraction. It combines a main path (with Depthwise Convolution and an SE module) and an interaction path (with 1 1 EConv and Projection Layers), synergizing convolutional operations with attention for richer multi-scale features.
Fig 4.
Architecture of the MultiSEAM module.
Integrated into the detection head, this mechanism tackles occlusion by leveraging depthwise separable convolutions, pointwise convolution for feature fusion, and a channel expansion module to integrate multi-scale context, thereby enhancing focus on informative features.
Fig 5.
Components of the SIoU loss function.
SIoU introduces direction-aware regression by combining Angle Cost (Λ), Distance Cost (Δ) weighted by angle, Shape Cost (Ω), and traditional IoU Cost (LIoU). This guides the model to a more efficient convergence path for superior localization accuracy.
Fig 6.
Example images from the BMR-Fall dataset.
The dataset provides benchmarks for challenging real-world conditions: (a) low-light conditions with degraded image quality, and (b) occlusion scenarios where the target is obscured by people or objects, ensuring robust model validation.
Table 1.
Hardware and software environment configuration for the experiment in this article.
Table 2.
The results are reported as mean ± standard deviation (mean ± std) from five independent runs. The p-values from the paired t-test and the 95% confidence interval (CI) for the performance differences (ΔmAP) between BMR-YOLO and each comparative model are shown in the table.
Table 3.
Algorithm comparison test results.
Table 4.
Comparison of ablation experiments.
Table 5.
Comparison of ablation experiments.
Fig 7.
Detection result comparison: YOLOv8nvsBMR-YOLO.
Qualitative results demonstrate BMR-YOLO’s superiority: (a) it reduces missed detections (false negatives); (b) it minimizes false alarms (false positives); (c) it provides more precise bounding boxes, confirming enhanced reliability in complex environments.
Table 6.
Benchmark results on the UR Fall dataset.
Table 7.
Generalization Validation on the Lei2 Fall Dataset.
Fig 8.
Performance comparison between YOLOv8n and BMR-YOLO models on the UR Fall dataset based on benchmarking results.
Fig 9.
Generalization performance of the BMR-YOLO model on the Lei2 Fall dataset based on generalization validation.