Robust Perception and Visual Understanding of Traffic Signs in the Wild
Rodolfo Valiente, Darren Chan, Alan R. Perry, Joshua Lampkins, Sasha Strelnikoff, Jiejun Xu, Alireza Esna Ashari
Abstract
As autonomous vehicles (AVs) become increasingly prevalent on the roads, their ability to accurately interpret and understand traffic signs is crucial for ensuring reliable navigation. While most previous research has focused on addressing specific aspects of the problem, such as sign detection and text extraction, the development of a comprehensive visual processing method for traffic sign understanding remains largely unexplored. In this work, we propose a robust and scalable traffic sign perception system that seamlessly integrates the essential sensor signal processing components, including sign detection, text extraction, and text recognition. Furthermore, we propose a novel method to estimate the sign relevance with respect to the ego vehicle, by computing the 3D orientation of the sign from the 2D image. This critical step enables AVs to prioritize the detected signs based on their relevance. We evaluate the effectiveness of our perception solution through extensive validation across various real and simulated datasets. This includes a novel dataset we created for sign relevance that features sign orientation. Our findings highlight the robustness of our approach and its potential to enhance the performance and reliability of AVs navigating complex road environments.