Physical activity and event detection with inertial sensors: Design and evaluation for low power microcontroller

Abstract

The aim of this thesis is to design inertial sensor based activity recognition for a low power continuous sensing architecture for today's mobile technology using a dedicated low-power sensor processor. The system framework is designed to provide higher level access to sensors and run sensor based algorithms. The sensor processor can provide consistently low latencies for feedback and interaction and continuous low power sensing for context adaptive user interfaces. In the current state of the art of hand-helds, every sensor is interfaced to the host processor directly and all the sensory data preprocessing and event detection is performed at the expense of the host processor. A dedicated sensor processor with good system architecture would improve the power and processing efficiency.

The objective of the thesis is to interface all inertial sensors to a low-power discrete sensor hub where this low power discrete sensor hub implements most of the sensor pre-processing and algorithm. This sensor hub is directly interfaced via Bluetooth link to the host processor which is an Android mobile device. The sensor firmware architecture is event driven and reduces the power consumption without the need for polling the sensors. Host processor sends control messages to the sensor box and the device when replies back to the control messages also produces sensor data so most of the traffic is from the sensor hub to the host.

Activity context detects user actions such as stand-still, walking, running, cycling, vehicle and sports. The activity context module uses naive bayesian classifier algorithm to classify the accelerometer sensor data and this module is implemented in the sensor box which has an ARM based microcontroller. Sensor drivers and sensor algorithms are tested using a test framework implemented in the application program. Event detection module implements free-fall detection in the sensor processor with the help of accelerometer data which shows the capability of the system. Classification algorithm verified in simulation environment shows around 92 percent accuracy in detecting the physical activity of the user.