Small Vehicle Damage Detection with Acceleration Spectrograms: An Autoencoder-Based Anomaly Detection Approach

Sara Khan

Abstract: This paper presents a Machine Learning (ML) based methodology for a real-time small-vehicle damage detection system. While dents and scratches may seem trivial for individual car owners, they hold significant implications for insurance companies and car-rental/taxi service providers. In addition to these interested parties, car manufacturers and car service providers are also keen to obtain accurate vehicle damage information for vehicle behaviour analysis. Therefore, vehicle damage detection remains a key and challenging problem in the automotive research community. Our novel approach mainly uses spectrograms of inertial sensors like accelerometers which measure accurate vehicle acceleration. Inertial sensors are cost-effective and tailored for real-time data capture which makes them a competitive candidate for our choice. The designed system employs auto-encoders as automatic feature extractors from input acceleration spectrograms. These feature representations derived from sensor data are then classified into damage or non-damage categories using an anomaly detection approach. We achieved an impressive approximately 90% F1-score. The outcomes of this work contribute to the progress of intelligent damage detection systems. Furthermore, it assists in understanding the relationships between auto-encoders and signal data in this context. Our approach can be used in many applications in the automotive sector, such as automated vehicle inspection, enhanced airbag response systems, and improving the safety of autonomous driving.