Real-time fall attitude detection algorithm based on iRMB

Aiming at addressing the practical needs of traditional fall detection algorithms that suffer from significant environmental interference, low average detection accuracy in complex scenes such as object occlusion, and high requirements for model inference speed, we propose a new high-accuracy fall posture detection algorithm, iRMB-YOLO, based on an improved YOLOv8 model with the iRMB module. Firstly, the optimized iRMB module is introduced into the backbone part of the model, performing convolution operations on each feature map to retain all information in the channel dimension, thereby enhancing the model's performance in processing low-resolution images. Secondly, the CReToNext module, after parameter tuning, replaces all C2f modules in the feature fusion layer, using aggregated cross-scale features for information interaction and integration of different scale features, further improving the model's performance without significantly increasing computational cost. Experiments demonstrate that on the public Fall Detection Dataset, iRMB-YOLO's mAP@0.5:0.95 improved from 88.4 to 95.1%, and the model's inference speed increased by 50.98%. Compared with other detection algorithms, iRMB-YOLO achieves higher average detection accuracy and inference speed with a slight increase in parameter and computation volume, meeting the deployment requirements of practical scenarios. For this study, we have open-sourced the project code at the following URL: 112345434/diedao (github.com).