Towards Safer Mixed Traffic: A Deep Learning Approach to Autonomous Vehicle Detection and Driver Awareness

Recently, many autonomous vehicle (AV) manufacturers have begun officially testing their vehicles on public roads. Companies such as Waymo, Cruise, and Zoox are actively trialing Level 3 and higher autonomy systems. AV deployment has already progressed beyond testing, with Waymo offering robotaxi services in San Francisco, Los Angeles, and Phoenix, and planning expansions to cities such as Austin, Atlanta, and Miami. While this progress is promising, it introduces a significant challenge known as the mixed traffic problem; where AVs and human-driven vehicles (HVs) must safely coexist on the same roads. Human drivers do not always exhibit ideal driving behavior, nor do they expect AVs to behave flawlessly. When AVs and HVs interact in close proximity, such as at intersections or highway merging points, the risk of collisions increases. To address this issue, we propose a methodology aimed at enhancing human driver awareness of nearby AVs and their behavior. This approach can be integrated into HV drivers' smartphones, advanced driver assistance systems (ADAS), or even deployed through infrastructure units at critical locations like intersections. As part of our work, we have compiled and publicly released an AV image dataset<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>https://github.com/melyabroudi/AutonomousVehiclesDataset. Using this dataset, we applied transfer learning on two lightweight classification models; You Only Look Once (YOLO) and MobileNetV2, to distinguish AVs from HVs. Our system delivers real-time advisories to HV drivers, helping them anticipate AV behavior and reduce crash risk. Among the models, YOLO achieved the highest accuracy, reaching 99.2% on the validation dataset.