Speaker
Description
Acoustic perception can complement vision-centric sensing in autonomous driving by providing safety-critical cues such as emergency-vehicle sirens and horns, which often include strong tonal components and distinctive modulation patterns. Acoustic sensing for autonomous vehicles spans event detection/classification, separation, localization, and tracking. This work focuses on sound source localization in urban scenarios. Recent approaches increasingly combine classical array processing with machine learning to improve robustness, but their effectiveness depends on the availability of large, labeled datasets. Model-based alternatives can account for propagation effects, yet high-fidelity urban acoustic modeling is computationally expensive and typically requires environmental information, including acoustic properties, that is rarely available in uncontrolled outdoor scenes. This paper investigates approximate acoustic modeling strategies for urban-like scenarios where the source may be hidden from direct view, for example around corners. We propose a propagation model based on virtual sources and evaluate the localization performance when combined with sequential measurements acquired along the vehicle trajectory. By exploiting motion, spatial diversity can be achieved over time, allowing pressure measurements to be collected at different locations sequentially and reducing the sensor count. To enable controlled evaluation, we employ a simulation-based framework to generate reference data for a moving receiver, capturing reflections, diffraction, Doppler effects, and absorbing boundary properties. The results indicate promising potential for acoustic propagation modeling in urban scenes to support localization under non-line-of-sight conditions, while keeping sensing requirements, model complexity, and computational complexity low.