Temporal Segmentation of Human Motion for Rehabilitation

Segmentation and Identification;Machine Learning;Optimal Control

Lin, Jonathan Feng-Shun

University of Waterloo

关键词: Segmentation and Identification;    Machine Learning;    Optimal Control;   

Others  :  https://uwspace.uwaterloo.ca/bitstream/10012/11764/3/Lin_Jonathan.pdf

瑞士|英语

来源: UWSPACE Waterloo Institutional Repository

【 摘 要 】

Current physiotherapy practice relies on visual observation of patient movement for assessment and diagnosis. Automation of motion monitoring has the potential to improve accuracy and reliability, and provide additional diagnostic insight to the clinician, improving treatment quality, and patient progress. To enable automated monitoring, assessment, and diagnosis, the movements of the patient must be temporally segmented from the continuous measurements. Temporal segmentation is the process of identifying the starting and ending locations of movement primitives in a time-series data sequence. Most segmentation algorithms require training data, but a priori knowledge of the patient;;s movement patterns may not be available, necessitating the use of healthy population data for training. However, healthy population movement data may not generalize well to rehabilitation patients due to large differences in motion characteristics between the two demographics. In this thesis, four key contributions will be elaborated to enable accurate segmentation of patient movement data during rehabilitation.The first key contribution is the creation of a segmentation framework to categorize and compare different segmentation algorithms considering segment definitions, data sources, application specific requirements, algorithm mechanics, and validation techniques. This framework provides a structure for considering the factors that must be incorporated when constructing a segmentation and identification algorithm. The framework enables systematic comparison of different segmentation algorithms, provides the means to examine the impact of each algorithm component, and allows for a systematic approach to determine the best algorithm for a given situation.The second key contribution is the development of an online and accurate motion segmentation algorithm based on a classification framework. The proposed algorithm transforms the segmentation task into a classification problem by modelling the segment edge point directly. Given this formulation, a variety of feature transformation, dimensionality reduction and classifier techniques were investigated on several healthy and patient datasets. With proper normalization, the segmentation algorithm can be trained using healthy participant data and obtain high quality segments on patient data. Inter-participant and inter-primitive variability were assessed on a dataset of 30 healthy participants and 44 rehabilitation participants, demonstrating the generalizability and utility of the proposed approach for rehabilitation settings. The proposed approach achieves a segmentation accuracy of 83-100%.The third key contribution is the investigation of feature set generalizability of the proposed method. Nearly all segmentation techniques developed previously use a single sensor modality. The proposed method was applied to joint angles, electromyogram, motion capture, and force plate data to investigate how the choice of modality impacts segmentation performance. With proper normalization, the proposed method was shown to work with various input sensor types and achieved high accuracy on all sensor modalities examined. The proposed approach achieves a segmentation accuracy of 72-97%.The fourth key contribution is the development of a new feature set based on hypotheses about the optimality of human motion trajectory generation. A common hypothesis in human motor control is that human movement is generated by optimizing with respect to a certain criterion and is task dependent. In this thesis, a method to segment human movement by detecting changes to the optimization criterion being used via inverse trajectory optimization is proposed. The control strategy employed by the motor system is hypothesized to be a weighted sum of basis cost functions, with the basis weights changing with changes to the motion objective(s). Continuous time series data of movement is processed using a sliding fixed width window, estimating the basis weights of each cost function for each window by minimizing the Karush-Kuhn-Tucker optimality conditions. The quality of the cost function recovery is verified by evaluating the residual. The successfully estimated basis weights are averaged together to create a set of time varying basis weights that describe the changing control strategy of the motion and can be used to segment the movement with simple thresholds. The proposed algorithm is first demonstrated on simulation data and then demonstrated on a dataset of human subjects performing a series of exercise tasks. The proposed approach achieves a segmentation accuracy of 74-88%.