UMNet: Human Fall Detection Via Universal Inverted Bottleneck and Multi-Scale Feature Fusion

Falls, occurring in various scenarios, can result in severe consequences. Fall detection systems enhance safety and rescue efficiency by promptly identifying fall events through realtime monitoring and analysis. The core of these systems is the accurate and rapid identification of falls via detection algorithms. In this paper, a visual fall detection network, UMNet, based on the Universal Inverted Bottleneck and Multi-scale Feature Fusion Module, is proposed. First, the lightweight hybrid backbone CUSNet is constructed using the CNN backbone and Universal Inverted Bottleneck to achieve better feature representation. Next, the three-level Multi-scale Feature Fusion Module is built employing the Scale Sequence Feature Fusion Module, Triple Feature Encoding Module, and C2f to effectively integrate semantic features at different resolutions. Finally, the lightweight detection head ADownHead is designed using ADown to ensure detection accuracy while reducing model size. Experimental results indicate that UMNet achieves a sensitivity of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 9. 4 \%}$</tex> and an AP@ 50 of 86.6 % on the custom dataset, representing improvements of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2. 3 2 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 4 1 \%}$</tex> over the baseline, respectively.