DapFall: Dynamic Amplitude Probability Density Profile based WiFi CSI Sensing for Fall Detection

WiFi channel state information (CSI) based fall detection is highly sensitive to different environments. Existing work ignores the CSI subcarrier mutual information which carries critical characteristic of each activity and is robust to environment. In this paper, we propose a data-efficient DapFall system, which enables cross-environment fall detection without the model retraining. The key insight of DapFall is to transform the CSI amplitude into dynamic amplitude probability density (DAPD), which is a 3D visual representation, strongly emphasizing the dynamic features. Further, DAPD retains subcarrier mutual information and meanwhile reduces static components. Then, we employ specific data augmentations to change the amplitude variation feature and further increase the diversity of fall dataset. DAPD enables an efficient cross-domain knowledge utilization using deep learning model R2+1D pre-trained in a totally unrelated task and then facilitate detection capability without the need for extensive data. Experimental results demonstrate that DapFall achieves an 94.3% recall (TPR) and 6.25% FPR in original environments with limited training samples and meanwhile exhibit robust performance in other environments without retraining the model.