| dc.description.abstract | Rollators/wheeled walkers are a commonly used mobility aid that is utilized by people with a wide array of mobility challenges such as Alzheimer ’s disease, stroke, multiple sclerosis, and general loss of stability and control associated with ageing or injury. The main benefit of walkers is that they are low cost, have a simple design, and allow users to maintain independence and continued physical activity. The largest concern with walker use is that accidents and improper use can cause falls. Falls can have serious repercussions such as broken bones, soft tissue damage, further loss of mobility, and increased fear.The focus of this research is a continuation of previous work focused on making walkers safer through understanding the use of assistive automation. Walkers with the addition of automation are termed Robotic Rollator (RR).A significant amount of research has been already done on RR. Two main strategies for RR research is either modifying existing manual rollators or creating a custom RR from the ground up. Each option has its difficulties. The modified rollators have the problem that commercial walkers have such a wide variation in designs that sensor readings may not be transferable from one model to another model. From experience on an earlier iteration of the WatkeRR project, force and odometry measurements can be significantly affected by flexing and sliding of the lightweight frames. For these reasons it was decided that a custom RR would be used. This prompted the development of the Waterloo RoboticRollator (WatkeRR). The WatkeRR was built to be a platform for measuring and testing low-cost controls and sensors. The base frame is designed to be able to match existing dimensions of most, if not all, commercial rollators by adjusting to match height, width, handle position, and wheelbase. It was important that the controls and data collection systems are untethered to allow for use outside a lab. To make the project more accessible, the controls are modular and built on an open source control structure.The open source modular nature of the platform allows for parallel projects to be conducted at the same time. The following features are discussed: force sensing handles, modular caster wheels, wheel force loading, odometry, top-level human/WatkeRR control, Robot Operating System (ROS) integration/simulation, vision systems, and machine learning. Controlling the human/walker system is a challenging problem as the system can vary even during use. Change in the control of a walker is different for each user, including gripping, fatigue, injury, or walker type. The critical challenge of the control problem is the fact that the desired path of the user is not known. The data from a user can also vary depending on their physical and mental state. To be able to determine the desired path without direct input from the user three separate approaches can be utilized: 1) forces applied to the walker through the handles, 2) tracking position/pose of the user, and 3) detecting probable path to a known object. While the scope of determining user intent being larger than this thesis, a major focus will be on object detection. Determining what objects of interest are around a rollator allows for an insight into user intent.The use of Convolutional Neural Networks (CNN), a deep learning approach that has shown promise for object detection is a field that has recently been advancing rapidly. This advancement has enabled safer and more effective autonomous vehicles. The second part of this thesis focuses on enabling object detection on the WatkeRR and integrating into a larger control strategy. The challenges of CNN application to the WatkeRR is that it needs to be achieved with fewer sensors, lower cost components, and less processing power.The navigation of WatkeRR can be performed indoors or outdoors which, due to highly variable lighting conditions, can cause problems with a vision system. As well, moving in crowds is not as much of a well-defined problem as moving on roads with a car.Only a few years ago, to train a CNN to detect even simple cases required days or weeks and access to a supercomputer. Now, top-level retraining on a commercial GPU can take a few minutes. This top-level training utilizes existing inception network structures. My work covers methods for obtaining training data, comparisons to a wide array of different input data as well as a hierarchical structure for improving precision. A significant amount of current work on CNN is focused on black box approaches that can determine a large number of classes. The disadvantage of this is that training takes significantly longer and detection accuracy for specific classes can be lower. With the improved training data and the application of the proposed hierarchical structure being used to train an established network. You Only Look Once (YOLO) used for detecting and localizing a wide range of objects.The thesis concludes with a set of recommendations on how to utilize this network in the ROS and how it will be integrated into path planning. | en |
