Auto++: Detecting Cars Using Embedded Microphones in Real-Time

In this work, we propose a system that detects approaching cars for smartphone users. In addition to detecting the presence of a vehicle, it can also estimate the vehicle’s driving direction, as well as count the number of cars around the user. We achieve these goals by processing the acoustic signal captured by microphones embedded in the user’s mobile phone. The largest challenge we faced involved addressing the fact that vehicular noise is predominantly due to tire-road friction, and therefore lacked strong (frequency) formant or temporal structure. Additionally, outdoor environments have complex acoustic noise characteristics, which are made worse when the signal is captured by non-professional grade microphones embedded in smartphones. We address these challenges by monitoring a new feature: maximal frequency component that crosses a threshold. We extract this feature with a blurred edge detector. Through detailed experiments, we found our system to be robust across different vehicles and environmental conditions, and thereby support unsupervised car detection and counting. We evaluated our system using audio tracks recorded from seven different models of cars, including SUVs, medium-sized sedans, compact cars, and electric cars. We also tested our system with the user walking in various outdoor environments including parking lots, campus roads, residential areas, and shopping centers. Our results show that we can accurately and robustly detect cars with low CPU and memory requirements.