Accurate and robust step detection and walking speed estimation algorithm for wrist-worn devices

Abstract

With the rise of interest in watch-type and wristband-type devices, wrist-worn devices have become a growing market in the field of wearable activity tracking devices. However, wrist-worn devices have clear disadvantages compared to existing waist-worn and ankle-worn devices including noise from various sensing conditions that cause irregular arm movements. This paper proposes two activity tracking algorithm for wrist-worn devices which overcome the suggested problems of wrist-worn devices. In chapter 1, I propose a step detection algorithm using three-axis accelerometer for wrist-worn devices. The algorithm consists of three phases, which address the problems of wrist-worn devices. The first data preprocessing phase calculates the Euclidean norm of the acceleration vector. It enables the algorithm to track the movement of a device only with the acceleration data. The second data filtering phase reduces the noise with a simple digital low-pass filter. Then, the third peak detection phase adopts a sign-of-slope method and average threshold method to accurately detect the step peaks under different sensor-carrying modes and speed conditions. A wrist-worn hardware prototype is designed and realized for algorithm evaluation. The experiment results show that the proposed algorithm is superior to the compared existing algorithm and commercial devices. The averaged detection error is approximately 1% in different test conditions. In chapter 2, I propose a robust walking speed estimation method with data from a six-axis inertial measurement unit (IMU), which is commonly mounted in wrist-worn devices, and user’s height information. The proposed method provides accurate walking speed estimation results under different sensor-carrying modes and walking speeds. The estimation is based on sensor-carrying mode detection with the TreeBagger model, and on Gaussian process regression (GPR) models, which are adapted to seven predetermined sensor-carrying modes. The speed estimation is done by calculating a weighted sum of multiple GPR models with the probabilities from TreeBagger model. To evaluate the superiority of my method, I implement it on a hardware. An experimental evaluation is performed on 16 healthy subjects with a treadmill. The experimental results show that the proposed method outperforms existing studies and comparable commercial devices for all sensing conditions. The averaged error of the proposed method is about 3% for all sensing conditions, while others show error of more than 15% in different sensing conditions. The results shows that proposed method has novelty compared to existing studies in terms of estimating walking speed accurately under changing sensing conditions only with single IMU sensor.