This paper aims to address two important aspects of human walking. First, detect the periodicity of gait, and second, detect the fall of a healthy individual. The fall is a common phenomenon in cluttered environments.
To address these two aspects, the data is collected through three-axis Inertial Measurement Unit (IMU) sensors for different walking patterns. The IMU-collected gait data are time series in nature, and gait is a cyclic process; hence, gait data consists of periodicity and repetitive patterns. It is very important to know the gait periodicity time of the elderly to know the risk of falling due to walking impairment. In this paper, an automatic gait cycle extraction technique is proposed based on the autocorrelation function of time series data. The assumption is that there is maximum energy (peak) in the acceleration signal during the initial foot contact of the heel strike, which is the point where one gait cycle finishes.
The mean stride duration was calculated for 15 subjects, which is 1.06 s. It is aligned with the bench-mark work reported for healthy individual gait cycle periods [0.98 s, 1.2 s]. The validation curve and graph are also presented. Another contribution is the detection of falls while performing different daily activities in the real world. For fall diagnostics, the IMU-based SiSfall data set is used, which includes two different health age groups: adults (18 years and above) and the elderly, for 14 different walking activities from different directions and magnitudes.
To automate identification, a hybrid deep learning model based on convolutional neural networks and long short-term memory (CNN-LSTM) is used to predict fall categories. The proposed CNN-LSTM model shows superior performance, with an accuracy of 99.12%. This research will provide the confidence to elderly individuals to walk independently.
The novelty of the work is the extraction of the gait cycle and the design of a personalized fall computation model with fast inference time suitable for real-time applications. The gait cycle prediction based on autocorrelation is significantly improved. To automate identification, a hybrid deep learning model based on CNN-LSTM is used to predict fall categories. The proposed CNN-LSTM model shows superior performance, with an accuracy of 99.12%. This research will provide the confidence to elderly individuals to walk independently.
