Determination of the geometry of the room by the impulse response using convolutional neural networks

Abstract

Determination of the geometry of the room by the impulse response using convolutional neural networks

Poleshchenko D.A., Melnikova K.E., Nazarov D.E.

Incoming article date: 06.10.2023

Existing methods for determining the geometry of an enclosed space using echolocation assume the presence of a large amount of additional equipment (sound sources and receivers) in the room. This paper investigates a method for determining the geometry of enclosed spaces using sound location. The method does not assume the presence of a priori knowledge about the surrounding space. One sound source and one sound receiver were used to create and capture real impulse characteristics. A microphone was used as a sound receiver and a finger snap was used as a sound source to produce the impulse response. In this work, we used convolutional neural networks that were trained on a large dataset consisting of 48000 impulse responses and a number of room geometry parameters corresponding to them. The trained convolutional neural network was tested on the recorded impulse responses of a real room and showed accuracy ranging from 92.2 to 98.7% in estimating room size from various parameters.

Keywords: convolutional neural networks, room geometry, echolocation, impulse response, robotics, recognition, contactless methods of measuring objects, sonar, geometry prediction, virtual reality