Automated Pose Estimation for the Assessment of Dynamic Knee Valgus and Risk of Knee Injury during the Single Leg Squat

Human Pose Estimation;Single Leg Squat;Inertial Measurement Unit;Classification;Risk of Knee Injury Assessment;Dynamic Knee Valgus Assessment

Kianifar, Rezvanaffiliation1:Faculty of Engineering ; advisor:Kulic, Dana ; Kulic, Dana ;

University of Waterloo

关键词: Single Leg Squat;    Dynamic Knee Valgus Assessment;    Master Thesis;    Inertial Measurement Unit;    Classification;    Risk of Knee Injury Assessment;    Human Pose Estimation;   

Others  :  https://uwspace.uwaterloo.ca/bitstream/10012/11779/3/Kianifar_Rezvan.pdf

瑞士|英语

来源: UWSPACE Waterloo Institutional Repository

【 摘 要 】

Many clinical assessment protocols rely on the evaluation of functional movement testssuch as the Single Leg Squat (SLS), which are often assessed visually. Visual assessmentis subjective and depends on the experience of the clinician. Developing a reliable automatichuman motion tracking and assessment system can improve the accuracy of SLSclinical assessments and provide objective results that can be tracked and monitored overtime to guide rehabilitation and determine an individual's response to an intervention.In this study, an Inertial Measurement Unit (IMU) based method for automated assessmentof squat quality is proposed to provide clinicians with a quantitative measure of SLSperformance.First, an automated pose estimation method is applied to SLS motion data. A setof three IMUs is used to estimate the joint angles, velocities and accelerations of thesquatting leg. To tackle noisy sensor measurements and gyro drift, a 7 degree of freedom(DOF) kinematic model of the lower leg was applied together with a constant accelerationassumption to approximate the angular velocity and linear acceleration at each sensorlocation. The kinematic model predictions of the angular velocity and linear accelerationand sensor measurements were fused via an Extended Kalman Filter (EKF). The position,velocity, and acceleration of each DOF were defined as the states to be estimated bythe EKF. The pose estimation results showed successful extraction of joint angles withan average RMS error of 3.2 degrees, 5.5 degrees, 7 degrees compared to joint angles estimated from motioncapture for the ankle, knee, and hip joints, respectively. For this estimation, the requiredparameters for the kinematic model, including information about the sensor placement andorientation as well as the kinematic link lengths, were extracted from the marker data.However, in clinical applications of the proposed method, when marker data is notavailable, these parameters need to be measured. Measuring these parameters is time consumingin the clinical setting, which limits application of IMUs for clinical purposes. Withthe motivation to make this procedure easier and faster, a method for approximating theparameters using placement assumptions and body measures was described. A sensitivityanalysis was performed to detect those parameters which most affect pose estimation accuracy.The sensitivity analysis results revealed that sensor orientation is the most criticalfactor for accurate pose estimation. In this thesis, a simple and easy to use method isproposed for sensor orientation calibration, based on a systematic placement of sensorsand using gyroscope information for orientation estimation. This protocol was evaluatedexperimentally and pose estimation error with approximated parameters before and afterapplying the calibration protocol were compared. The comparison results showed that theestimate of the sensor orientation increases the pose estimation accuracy by 6.5 degrees for theknee joint angle and with an average of 1.8 degrees for other joints without the need for timeconsuming calibration.In the second part of the thesis, an algorithm for automated assessment of the SLSin terms of dynamic knee valgus and risk of knee injury is developed. After applying thepose estimation algorithm to IMU data of SLS motions, the estimated time series dataof joint angles, velocities and accelerations for consecutive squats were segmented intoindividual squat repetitions. Statistical time domain features were generated from eachrepetition. The most informative features were selected using a combination of 18 featureselection techniques. Six common classifiers in including SVM, Linear Multinomial LogisticRegression, Decision Tree, Naive Bayes, K Nearest Neighborhood, and Random Forestswere applied to the full dimensional data, the subset of selected features, and extractedfeatures by supervised principal component analysis. The proposed approach was evaluatedin two trials. First, a pilot study was conducted on a small dataset, followed by analysison a larger clinical data set, collected by our clinical collaborator. For the clinical study, adataset of SLS performed by healthy participants was collected and labelled by three expertclinical raters using two different labeling criteria: ;;observed amount of knee valgus;; and;;overall risk of injury;;. Labels included ;;good;;, ;;moderate;;, and ;;poor;; squat qualityor ;;high risk;;, ;;mild to moderate risk;;, and ;;no risk;; of injury. Feature selection resultsshowed that both flexion at the hip and knee, as well as hip and ankle internal rotationare discriminative features, and that participants with ;;poor;; squats bend the hip andknee less than those with better squat performance. Furthermore, improved classifi cationperformance was achieved by training separate classifi ers strati ed by gender. Classifi cationresults showed excellent accuracy, 93.1% for classifying squat quality as ;;poor;; or ;;good;;and 95.3% for differentiating between high and no risk of injury.