Preparing the Automated Future for Wheelchair Users Through Motion Prediction and User Input-Based Intent Inference

Open Access
- Author:
- Wolkowicz, Kelilah Louise
- Graduate Program:
- Mechanical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- November 15, 2018
- Committee Members:
- Sean N Brennan, Dissertation Advisor/Co-Advisor
Sean N Brennan, Committee Chair/Co-Chair
Jason Zachary Moore, Committee Member
Henry Joseph Sommer III, Committee Member
Bruce Gluckman, Outside Member - Keywords:
- Autonomous vehicles
Band-pass filters
Bar codes
Biopotentials
Brain-computer interfaces
Computerized instrumentation
Data acquisition
Design automation
Detection algorithms
Discrete transforms
Electroencephalography
Electromyography
Electrooculography
Estimation theory
Gaze tracking
Human computer interaction
Embedded systems
Indoor localization
Input variables
Intelligent vehicles
Kalman filters
Linear discriminant analysis
Linear feedback control systems
Linear systems
Mechanical engineering
Mechanical Engineering
Mechanical systems
Microprogramming
Motion control
Odometry
Path planning
Pose estimation
Prediction algorithms
Probability
Probability distribution
Programming
Rehabilitation robotics
Research and development
Robot control
Robot kinematics
Robot motion
Robot programming
Robotic equipment testing
Simulation
Simultaneous localization and mapping
State estimation
State feedback
System implementation
System integration
System validation
User interfaces
Virtual reality
Wearable sensors
Biopotential signals
Human-computer interfaces
Decision-making
Intent inference
Slip detection
Motion prediction
Robotic wheelchair - Abstract:
- This dissertation focuses on developing algorithms, methods, and technologies to further enhance and promote wheelchair patient independence. The research is motivated by the limitations wheelchair users face and the need to produce safer assistive devices that require reduced user input. The work in this dissertation ranges from algorithm development to experimental implementation on a real-world robotic wheelchair system. Specifically, the challenge of improving the performance of wheelchair guidance, even in the presence of challenging conditions and/or noisy, ambiguous sensory inputs from the user, is addressed through three complementary objectives. In the first objective, control and estimation algorithms are developed that have the ability to predict wheelchair motion for changing environments, users, and wheelchair designs within 10% error. A robust, fast-dynamic, inner control loop is designed to encompass the plant, including inputs to a robotic wheelchair and raw sensor outputs, sensing through sensor fusion, disturbances, and a controller for PID-based local control. The results demonstrate the application of an instantaneous center of rotation (ICR) estimation for a wheelchair under situations where ICR locations should not typically change. This method estimates wheelchair tire slip in real-time and predicts motion during that slip using a model-based framework. Experiments show that the ICR locations do not vary significantly under the slip-free conditions of normal operation, with 2-sigma standard deviations of 0.076 m, but the ICR locations deviate up to 0.84 m during slippage. Additionally, wheelchair motion is predicted within an error of 0.11 m in terms of Euclidean distance over 6 m total Euclidean distance traveled, compared to an error of 1.22 m over the same 6 m total Euclidean distance traveled when localizing the wheelchair using wheel odometry alone. While this objective enhances wheelchair user safety, user intent is still difficult to infer. Thus, the second objective is to infer paths for wheelchair guidance within 10% error, using both spatial and temporal wheelchair path information. The methodologies applied in this objective reduce user joystick inputs by 50 to 70%, as well as reduce path tracking error during wheelchair navigation and guidance. Results demonstrate that wheelchair trajectories can be algorithmically grouped into similar paths to help discern the probability of a particular goal destination along a path. Moreover, it was found that blending joystick inputs with both spatial and temporal path information enables better inference of path-making decisions and wheelchair navigational guidance. The implementation of this technique in a real-time experimental scenario resulted in an average reduction of 73% of required user inputs. Finally, the third objective of this dissertation is to determine whether path decision-making is aided by integrating user biopotential signals, such as EEG, EOG, and EMG, with joystick inputs. The particular challenge of this objective lies in the reliability and repeatability of biopotential signals. Moreover, the measurement of biopotential signals on a moving robotic wheelchair have been found to be corrupted by motion artifacts. Testing shows that while EOG and EMG inputs can be blended with joystick inputs to improve the identification of user intent during wheelchair navigation, preliminary EEG measurements of motor imagery are not easily classified, therefore, further work is required. In conclusion, this dissertation develops a model-based framework that predicts wheelchair motion for changing environments within 3% error. Using both spatial and temporal wheelchair path information, user joystick inputs are reduced by 73%. When incorporating both EOG and EMG user biopotential signals with joystick inputs, the path decision-making process results in an 85% accuracy rate.