An indoor fall detection system based on WiFi signals and genetic algorithm optimized random forest

• Published in: Wireless networks 2024-04, Vol.30 (3), p.1753-1771
• Main Authors: He, Jiai, Zhu, Weijia, Qiu, Lili, Zhang, Qin, Wang, Chanfei
• Format: Article
• Language: English
• Subjects: Background noise; Communications Engineering; Computer Communication Networks; Discrete Wavelet Transform; Electrical Engineering; Engineering; Fall detection; Genetic algorithms; Human activity recognition; Human influences; IT in Business; Networks; Original Paper; Performance enhancement; Phase shift; Subcarriers; Variance; Wavelet transforms
• ISSN: 1022-0038; 1572-8196
• DOI: 10.1007/s11276-023-03625-w

Human activity recognition technology has received much academic attention in recent years and plays a vital role in a wide range of applications such as smart healthcare and security surveillance. A fall detection scheme based on Channel State Information (CSI) is designed to overcome the limitations of traditional fall detection systems such as low privacy, high cost and poor cross-domain capability. Based on the data collected by commercial WiFi devices, the scheme analyzes the changes of indoor CSI caused by human activities and realizes the non-contact fall detection. Firstly, Discrete Wavelet Transform (DWT), phase difference analysis and Moving Average Filter (MAF) are combined to reduce the impact of environmental noise on the detection performance. Secondly, the variance of CSI amplitude and phase difference is calculated as an indicator for selecting subcarriers and the moving variance is used to segment the active interval of the selected subcarriers, which can reduce the data dimensionality and extract the time-frequency features. Finally, the Genetic Algorithm (GA) is used to optimize the parameter selection process of the Random Forest (RF) to improve the performance of the classifier model. Through comparison experiments on the DARMS dataset, it is verified that the proposed scheme can effectively improve the accuracy of fall activity recognition and maintain an accuracy of over 95.25 % when the training environment changes.