Wearable Inertial Sensors for Motion Analysis in Respiration, Diet Monitoring, and Vehicular Safety Applications

Abstract

This thesis is concerned with the development and application of motion analysis algorithms based on signals from inertial measurement units (IMUs). In particular, the application areas discussed in the thesis are respiratory monitoring, dietary monitoring, and vehicular safety. Usage of IMUs for attitude and heading estimation has a rich legacy, but it is only in recent years that they have become low-cost commodity sensors found in nearly every smart phone and smart watch, making them particularly applicable sensors for everyday applications. Despite the existence of well-established orientation estimation techniques, motion analysis using inexpensive wearable sensor applications targeted to the general population requires special attention. All three application areas discussed in this thesis require a similar approach to the estimation of motion variables in that they depend on the partial or full orientation of the device relative to the human user and/or the user’s orientation relative to earth. However, the class of mobile-phone grade IMUs utilized here offer notoriously poor accuracy compared to much more expensive aerospace-grade IMUs. Inexpensive IMUs typically suffer from bias instability, which requires careful calibration or specialized algorithms. Further, full orientation estimation traditionally relies on the IMU’s magnetometer to sense the geomagnetic field. But, the geomagnetic field is relatively weak and can often be dwarfed by magnetic fields from ferromagnetic objects routinely encountered in indoor environments. Thus, applications targeted for use by the general population must utilize algorithms that can overcome these limitations in a robust manner. The first application area addressed is respiratory monitoring. The physical motions of the thoracoabdominal wall during respiration are important in many diseases, and differentiation of normal from abnormal respiratory kinematics can be used to monitor disease state. In this application, a novel wearable device is developed that allows for long-term, out-of-clinic monitoring and analysis of respiration under the assumption of static body position (non- ambulatory). In particular, the device measures respiratory accelerations at multiple points on the thoracoabdominal surface and estimates respiratory displacements along with a variety of clinically useful metrics. After careful removal of gravity from the acceleration measurement using a multiplicative Kalman Smoother, the algorithm double integrates and high pass filters the residual signal to obtain three-dimensional respiratory displacements. The accuracy is on the order of the accuracy of a reference optical motion tracking system, and this thesis presents an analysis of the factors contributing to displacement errors. From the displacements, a variety of additional temporal, phasic, and volumetric respiratory variables may be estimated. After developing methods and discussing experimental results from a single subject for estimating respiratory displacements and subsequently several respiratory variables from these displacements, we then present the results from an initial small cohort IRB-approved study using the device. In the study, subjects wore the respiratory monitor while faced with a variety of airway occlusions. Despite the ultra-low respiratory rates encountered, the system was able to detect thoracoabdominal asynchrony with limited accuracy. Real-life medical situations involving respiratory distress are likely to present higher respiratory rates and thus higher potential for more accurate estimates. The developed system offers a combination of capabilities unmatched by existing technology in terms of its portability and the suite of respiratory variables it is able to estimate. The second application area addressed in this thesis is the development of a novel Food Intake Monitoring (FIM) device. Typical methods for dietary tracking in obesity research such as the 24-hour food intake recall are well known to be inaccurate, and there is clear need for a device to automatically detect and capture eating events as an adjunct to these existing methods. In this thesis, a wrist-worn IMU and microcontroller are utilized to detect when a person is eating (under the assumption that the food is eaten primarily with the sensor-affixed arm), optimize the capture of the food being eaten using an on-board camera, and classify the obtained image as containing food or not. The detection, image capture, and classification modules are organized in a decision tree format, an approach which minimizes system power consumption while maximizing user privacy, as opposed to having a camera always on with constant wireless data being streamed. In the first iteration of the FIM, hand proximity to the mouth is decided based on two IMUs, one on the upper arm, and one on the lower arm. In the second iteration of the device, only a single IMU is utilized, and hand proximity is determined using the IMU’s magnetometer along with a magnet worn on the body near the collar bone. Once hand-mouth proximity has been detected, it is shown that a simple linear Support Vector Machine is able to accurately classify eating activities versus other hand-near-mouth activities, such as teeth brushing and shaving. After eating is detected, the system takes an image of the food in front of the user using an on- board camera. The timing of the image capture is based on estimation of the device orientation relative to gravity using a straightforward Kalman Filter, and a method is developed that predicts optimal image capture timing using the gyroscope. Finally, it is shown that images may be classified as containing food or not using a special Convolutional Neural Network (CNN) adapted to microcontroller deployment using integer quantization. The final health and safety application considered concerns vehicular safety and phone use while driving. Distracted driving due to phone or mobile device usage is one of the primary causes of vehicular accidents, and one approach to reducing such accidents is to automatically disable devices when the user is driving. In this thesis, IMU signals on a mobile phone or smart watch are utilized to determine whether or not the user is in the driver’s seat of a moving vehicle, under the assumption that the device is in a static position inside the vehicle and close to level road grade. First, the algorithm must estimate the orientation of the device relative to the vehicle. As in the other applications, fundamental limitations of mobile-phone grade IMUs prevent estimation of orientation using traditional methods. Instead, the algorithm uses motion signals obtained during braking to determine the forward direction of the vehicle, while estimation of the gravity direction fully constrains the phone orientation. Once the orientation is determined, the pitch and roll dynamics encountered during braking and turning the vehicle are used to determine which quadrant of the vehicle the device is in relative to the vehicle’s center of gravity. Successful identification of seat position is demonstrated first in simulation and then experimentally using data taken during real-world city driving conditions.