A pathological tremor is an involuntary and periodic motion of a body part. The detection and quantification of a pathological tremor are essential for diagnosis and therapy. The goal of this research is to detect the frequency of the pathological tremor. Instead of detecting tremors by using a specific medical device, we propose a new architecture jointly using a state-of-the-art pose estimation method and periodicity detection technology to identify pathological tremors from a video. In our approach, an advanced deep neural network is deployed for human pose estimation. A pixel-wise method for frequency estimation is designed to spatially integrate the spectral information of pixels to refine an estimate.

Compared with conventional methods, our method offers significant convenience for both patients and medical staff. Our approach does not need a specific device. Thus it eliminates the error caused by the additional mass of the sensor. The procedure of the test is simple for non-technical staff so that the method decreases the possible operational error.

Each module is evaluated on a real dataset by a series of experiments. Compared with a classic 1D surrogate signal method, our pixel-wise method has a smaller error and deviation on both synthetic videos and real videos. The architecture is finally evaluated on patient videos and shows a promising result. For 21 periodic videos, 13 of our frequency estimations have an absolute error lower than 1 Hz.