Abstract:

A scenario robust fall detection system based on few-shot learning (FDFL) in wireless environment was designed.The performance of existing fall detection methods based on Wi-Fi channel state information (CSI) degrades significantly across scenarios, which requires collecting and marking a large number of CSI samples in each application scenario, resulting in high cost for large-scale deployment.Therefore, the method of few-shot learning was introduced, which can maintain the performance of fall detection with high accuracy when the number of annotated samples in unfa-miliar scenes is insufficient.The proposed FDFL was mainly divided into two stages, source domain meta-training and target domain meta-learning.The meta training stage of the source domain consists of two parts: data preprocessing and classification training.In the data preprocessing stage, the collected original CSI amplitude and phase data were denoised and segmented.In the classification training stage, a large number of processed source domain data samples were used to train a CSI feature extractor based on convolutional neural network.In the meta-learning stage of the target domain, the limited labeled data sampled in the target domain was effectively extracted based on the feature extractor trained in the meta-training module, and then a lightweight machine learning classifier was trained to detect the fall behavior under the cross-scene.Through several experiments in different scenarios, FDFL can achieve an average accuracy of 95.52% for the four classification tasks of falling, sitting, walking and sit down with only a small number of samples in the target domain, and maintain robust detection accuracy for changes in test environment, personnel target and equipment location.

Key words: Wi-Fi sensing, fall detection, CSI, cross-domain detection, few-shot learning