Car accident detection and reconstruction through sound analysis with crashzam

Abstract

A car accident is a dangerous and extremely stressful moment involving many actors: one or more drivers, passengers, rescues, police force, as long as repairmen and insurance companies. In such circumstances, smart connected vehicles have the opportunity to provide really helpful services like an automatic crash detector for quick first aid alerting, first notification of loss, proactive claim management, settlement estimation, and damage assessment. While until now crash detection and reconstruction is completely based on accelerometer sensor time series analysis, this paper presents an innovative way to detect any type car accidents, called Crashzam, which relies on another source of information: the sound produced by a car impact that reverberates inside the car cabin. In the manuscript we introduce an original dataset we have beforehand built composed by in vitro and in vivo crash sounds, the overall system design, model and classification built upon features extracted from the time and frequency domain of the audio signal and from its spectrogram image, and a possible approach for crash reconstruction. Results show that the proposed model is able to easily identify crash sounds from other sounds reproduced in-car cabins, even in presence of noise.