A real-time hidden Markov model based action recognition system using body sensor networks.

摘要

We describe a body sensor system that detects human activities in real-time. The system consists of wearable computers known as sensor nodes (motes) that can sense information, process them and transmit the results to a Personal Device like Smart phone, PDA or Personal Computer. The motes are attached to different parts of the human body, namely waist and right thigh. To conserve power on the sensor motes, we send labels rather than the actual sensor data. The labels are a small set of alphabets that can completely represent the sensor data are generated by running Gaussian Mixture Model (GMM) clustering algorithm. The labels sent by the motes are analyzed by a HMM classifier to detect various actions performed by the subject. We will describe how various signal processing and statistical algorithms were adapted to run efficiently on the motes as well as on the PC, with minimal loss of precision. Motes lack floating point operations, therefore, signal processing techniques with floating point precision need to be adapted to execute with fixed-point operations. We also report the overall performance of the system in terms processing time for various modules and the action recognition accuracy. The thesis also explores mechanism to improve the robustness of the system by rejecting uninteresting movements.