Movement Classification and Analysis from RGB – D Video Data
MetadataShow full metadata
The aim of this thesis is to develop and evaluate methods of human movement classification using motion tracking data captured using a RGB-D sensor. As hardware solutions evolve and improve, so too must the software solutions evolve to creatively leverage and combine new technologies. Accurate human movement classification can facilitate a variety of practical applications, ranging from the health domain to sports, film, and even advanced surveillance. In this work, we focus on human movements related to physical therapy exercises. Motion tracking data was collected from subjects performing various physical exercises. The goal of our system is to automatically recognize the types of exercises performed by teach subject and the number of repetitions of each particular exercise. To achieve this goal, we use 3D skeleton tracking data points provided by the Microsoft Kinect sensor. After a set of transformation steps, we apply a Long Term Short Term Memory (LSTM) Deep Learning networks in tandem with the Dynamic Time Warping sequence matching algorithm to classify the types of exercises and number of repetitions performed by each subject.