Application of Signed Distance Function Neural Network in Real-Time Feet Tracking

Loss of locomotion is one major problem faced by the elderly currently. Various rehabilitation technologies have been developed to assist them in recovering walking capability. Gait monitoring is an important aspect of lower-limb function rehabilitation. By observing the walking behaviour of a patient, the stability and recovery progress can be evaluated. Despite gait is often measured by motion capture and force-based measurement, these methods are costly and non-portable, whereas inertial measurement units (IMUs) require the attachment of sensors onto the subject's body. A few contactless measurements have been proposed, however, none of them views the feet from the back, making it non-trivial to transfer the method to over ground rehabilitation robots. This paper proposes a method to track the poses of the feet in real time using a novel deep neural network, termed SDF-Net, that models the signed distance function (SDF) of an object. Independent of the viewing angle, the algorithm receives the colour and depth images of the feet as input and computes the pose of the feet. The tracking accuracy is evaluated by having a subject to perform various actions with the feet; the dynamic errors are found to be less than 9 mm and 8 degrees for position and orientation errors respectively, which are better than the state-of-arts reviewed.