Object Position Labelling in Video Using PRBS Audio Multilateration
Pádraic McEvoy · Arrow - TU Dublin (Technological University Dublin) · 2019
Supervised machine learning approaches for tracking objects’ positions in video typically require a large set of images in which the positions are labelled. Human labelling is time-consuming and automatic position labelling using visual markers is generally not possible because visible markers would corrupt the data. Here, we present an approach in which an object is tracked using a hidden tag that emits a PRBS audio signal. Four microphones arranged in a planar cross formation capture parallel recordings of the PRBS signal. Multilateration, using the time difference of arrival (TDoA) of the PRBS at each microphone, is used to estimate the position of the emitter. Here, we describe and evaluate the method by which the TDoAs are obtained and the emitter position is calculated. When evaluated, the approach yielded threedimensional position estimates with a mean error of 18.56cm. In its present form, the method is suitable for applications in which precision is not a priority, but three-dimensional object coordinates are required rather than two-dimensional camera view coordinates.