Patient Ambulation Assessment Using Depth Cameras

In the field of medical care, effective assessment of human activities is essential for understanding the patient's physical state in general and postoperative recovery, in particular. Detecting human joints by using depth cameras and computer vision methods is an important means of assessing human postures and activities. This thesis explores methods of evaluating human activity from three angles by using the second generation of Microsoft Kinect camera. The first part focuses on quantifying human activity using three metrics: average speed, distance traveled, and postures. In particular, by tracking the human head, its position relative to the camera can be determined with reasonable accuracy. By setting thresholds for different postures, the posture of the human body can also be determined. Accuracy of at least 87% was achieved for distance traveled and average velocity measurements. For posture detection, accuracy of at least 80% was achieved. The second part demonstrates a subject identity recognition method by measuring the height of the targeted human body and the distances among their joints. The distances between adjacent joints and height of a subject’s head are used to create a vector of eight features for an individual to use for identification. Using a modified KNN, full and partial feature sets were used to identify subjects. The classification results were promising, and the mean accuracy for all subjects reached 95.3%. In the third part, we proposed a method for posture detection based on tracking part of human joints. To differentiate static postures and dynamic movements, a hierarchical classifier is used. By analyzing the relative positions of the tracked joints, key features can be extracted as the basis for static posture classification. In addition, two indicators of speed and acceleration can be used to identify dynamic postures. In the specific classification stage, we used the Support Vector Machine (SVM) method. The performance of SVM shows that the average accuracy of the entire hierarchical classifier is 97.86%.