Methodology of activity recognition: features and learning methods

Abstract

Sensor-based Human Activity Recognition (HAR) has been explored by many research communities and industries for various applications. Conventional pattern recognition approaches based on handcrafted features contributed a lot in this research field by employing general classification approaches. This chapter represents those handcrafted features in time and frequency domain along with their importance and feature selection methods. Following these methods, this chapter provides explanations of several conventional machine learning techniques for classifying sensor data for activity recognition. The problems of overfitting and underfitting have been discussed with remedies. Previous research works using conventional pattern recognition (PR) approaches on some benchmark datasets have also been analyzed.