Meaningful Audio Data Augmentation for Approaching Vehicles Detection

Road safety is of extreme importance worldwide, affecting millions of lives annually. To solve the problem of road accidents, passive and active countermeasures have been adopted; among passive solutions, vehicle detection and driver alerts are becoming more common. Audio signal analysis as a possible solution to extract an incoming vehicle and its speed is a current research topic, and many authors have tried to analyse audio for vehicle detection using different techniques. Artificial intelligence and neural networks are promising approaches; however, they present the problem of creating audio datasets. Finding controllable environment noises and acquiring audio signals of different vehicles at increasing speed requires a massive amount of time and special design facilities, both challenging to obtain. Data augmentation in machine learning is commonly adopted to solve this problem and expand an already existing dataset; in the context of audio signals, the numerous developed and available techniques are mainly focused on speech and speaker recognition. In this work, we present a possible solution for audio signal augmentation based on the introduction of environmental noises, such as rain and wind noises, to construct a vehicle detection system based on audio classification. We present preliminary results of audio signal augmentation and explain how the consciously chosen addition of noise can improve a driveraltering system based on audio signal classification.