Elderly fall classification using passive Wi-Fi radar and machine learning

In the context of Malaysia's changing demographic landscape, the increasing elderly population necessitates innovative solutions for their healthcare and safety. Fall incidents among the elderly significantly affect their health, prompting a focus on assistive living technologies aimed at fall detection. This study aims to enhance fall detection by developing a system using Passive Wi-Fi Radar, addressing limitations of previous approaches. This method alleviates issues related to the comfort of wearable devices, privacy concerns of vision-based devices, and detection limitations of other ambient sensors. Passive Wi-Fi Radar, initially intended for communication, presents challenges in detecting human falls and movements. To address this, an enveloping algorithm is introduced to mitigate the influence of beacon signals carrying trajectory information. Additionally, four signal transformations which are Fourier spectrum, power spectrum, scalogram, and spectrogram are employed to analyse and extract Amplitude modulation signatures related to falls and movements. Feature extraction and selection are then applied to identify points that distinguish various fall types and fall-like movements. For Fourier and power spectrum features, extraction and selection are performed using Principal Component Analysis, followed by classification tasks utilizing Support Vector Machine, Random Forest, and Feed Forward Neural Network. Spectrogram and scalogram features are fed into the AlexNet Convolutional Neural Network. This study also explores combining AlexNet feature layers with traditional machine learning techniques to create hybrid classifiers, improving the performance of traditional algorithms. Results show that the Hybrid Random Forest classifier, using spectrogram images, performs exceptionally in single person scenarios, achieving an accuracy of 97.5%, recall of 0.9684, precision of 0.9872, and an F1-score of 0.981. In multiple person scenarios, the classifier attains an accuracy of 98.51%, recall of 0.9786, precision of 0.9914, and an F1-score of 0.9888. These findings highlight the effectiveness of the proposed fall detection system in addressing the challenges posed by the elderly population, marking a significant advancement in assistive living technologies. In conclusion, this study demonstrates the feasibility of using Passive WiFi radar to detect falls, even high-speed falls. The introduction of hybrid classifiers has notably enhanced the model's overall performance, representing a valuable improvement for utilizing Passive Wi-Fi radar in fall detection. This research provides important insights and advancements in fall detection technologies, paving the way for more effective and accurate solutions to address the healthcare and safety concerns of the growing elderly population.