New method for human activity recognition based on IMU sensors and digital speech processing theory

Abstract

This paper presents a new method for human activity recognition (HAR). Nowadays the biggest parts of HAR systems are relying on wearable IMU (inertial measurement unit) sensors. The common IMU sensors are accelerometers and gyroscopes. These sensors are widespread in mobile devices such as smartphones and smart watches. Authors usually use real time signal features as inputs for classifiers that are calculated using sliding windows only. This paper proposes a new method based on speech-silence discrimination technique for detecting the beginning and the end of an activity. The presented method relies on short-time log energy (STLE) and cumulative sum of angle of STLE values. The method was tested on two similar physical activities: squat and knee raise. This algorithm provides a 41.2% pre-classification accuracy, by precise detection of the length of individual exercise states (start, intermediate, and finish position) only. The proposed method reduces complexity, classifying only activities when they are detected (not classifying pause between activities).

Keywords

• human activity recognition
• short-time log energy
• cumulative sum
• activity state detection
• IMU sensors
• physical exercises