An Indoor Area Estimation Method Analyzing Spectrograms of Environmental Ultrasounds by Convolutional Neural Network

摘要

We propose and evaluate a method for estimating an indoor area by using spectrograms of environmental ultrasound. The method does not require special transmission devices to be installed beforehand. The method uses a convolutional neural network (CNN) whose inputs are spectrograms of ultrasounds. We collected ultrasounds at frequencies of up to 125 kHz in 10 rooms at a university, and evaluated the proposed method based on the collected data. The area classification accuracy was 0.978, showing that the proposed method is effective. An experimental result also suggested the proposed method has robustness against changes over time when learning data is sufficiently collected on multiple days. An evaluation on the influence of recording positions in a room showed that classification accuracy is influenced by a distance between recording positions. Moreover, an experimental evaluation showed that a band of 33.125 to 72.5 kHz is the most important in our experiment.