A pervasive and sensor-free Deep Learning system for Parkinsonian gait analysis

Parkinsonian gait is associated with life-threatening consequences such as fall risk in Parkinson patients. Conventional Parkinsonian gait analysis heavily relies on expensive sensors and human labor. In this work, we propose a sensor-free end-to-end system which enables the automated and accurate Parkinsonian gait detection and analysis upon the videos recorded by pervasive cameras. Specifically, we leverage Deep Learning technologies to extract the human skeleton in the video frame and address the camera random angle challenge. By analyzing the gait features, we train a classifier based on a binary decision tree. Out of 16 Parkinsonian gait and 13 healthy gait videos, our system is able to detect the Parkinsonian Gait with 93.75% accuracy and healthy gait with 100% accuracy.