AUDIO-SIGNALS CLASSIFICATION METHODOLOGY USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

The development of telecommunications, the Internet, social networks, and various mobile devices for recording audio and video information leads to an explosive growth in the number of multimedia data, in particular, audio files. An example of sound data is recordings of sounds of various origins, videos from surveillance cameras and video recorders, mobile phones, etc. That is why there is a need to automate the search, recognition, identification and classification of audio-visual information.

This article describes the methodology for identifying and classifying car accidents based on their audio signals and the mathematical apparatus of convolutional neural networks. The main task of the developed methodology is to improve the accuracy of accident classification and improve the energy efficiency of the methods and algorithms used for this purpose in order to ensure their operation in real time on mobile and peripheral devices with limited computing power. The relevance of this task is driven by new car safety standards and the introduction by the countries of the European Union and the United States of requirements for equipping all new cars with automatic accident notification systems (Eng. Automatic Crash Notification, ACN). Since the response of emergency services (police, ambulance, etc.) to such reports is extremely expensive, the task of increasing the accuracy of methods of analyzing road events arises.

Nowadays, most automakers use accelerometer sensors for this purpose. However, the use of acceleration data alone often leads to misclassifications: pits on the road are defined as an accident, while a rear impact can be classified as normal acceleration. As a result, the average accuracy of such techniques does not exceed 85%, which is acceptable, but obviously requires further improvement. In recent years, a significant number of modern studies have appeared that propose the use of convolutional neural networks for classifying audio signals of emergency events. The effectiveness of such approaches is significantly higher and is at the level of 85%-95%. The best results of 96% were achieved by authors [9], who used an ensemble of three convolutional neural networks. Nevertheless, this approach requires considerable energy and cannot be applied for real-time accident detection on peripheral devices with restricted processing power.

This article describes the methodology of car crash classification based on the crash audio signals and the use of convolutional neural networks as a classifier, achieving higher accuracy of 98% and better energy-efficiency comparing with other existing methodologies