Sound Source Localization of Cars at Intersections Based on Deep Learning

Highly automated and autonomous vehicles currently rely on three complementary main sensors to identify visible objects, namely cameras, lidar and radar. However, the function of these traditional sensors may be limited in urban environments when the sight is blocked by narrow streets, trees, parked vehicles and other traffic. Acoustic perception does not depend on the line of sight, Multiple, cheap microphones are used to capture sound as an auxiliary sensing mode for the early detection of vehicles approaching behind blind spots in an urban environment. We use a CNN to learn on data converted to Mel spectrum features and achieve 86% accuracy, our program can determine the direction of approaching cars at any traffic intersection.