This study proposes a lightweight deep learning-based fall detection model that effectively determines fall events using multi-sensor signals collected from different body locations. The proposed model incorporates dilated causal convolutional blocks with various kernel sizes, achieving both model compression and improved accuracy. To validate its performance, we utilized the publicly available UMAFall dataset, which includes multi-sensor signals collected from the chest, waist, wrist, and ankle. The proposed fall detection model was compared with LRN, a lightweight variant of the ResNet18 architecture. Experimental results demonstrate that the proposed model is approximately four times lighter than LRN, while consistently achieving superior performance across all sensor locations. In particular, our model achieves macro-averaged F1-scores of 98.31% (waist), 98.06% (chest), 95.68% (ankle), and 93.54% (wrist), outperforming LRN by 0.65%, 1.54%, 2.22%, and 3.87%, respectively, despite its significantly reduced model complexity. These results underscore the model’s potential for accurate and efficient fall detection in practical applications involving wearable sensor systems.
