Traffic sign 3D recognition: A monocular approach based on keypoint matching

Traffic Sign Detection and Recognition (TSDR) is a critical component of many automotive applications, such as life-saving driver assistance. Accurate 3D recognition of traffic signs is indispensable for a correct interpretation of a sign’s content. Despite its importance, the topic of traffic signs’ 3D recognition has not received adequate attention from the community. In addition to their scarcity, existing solutions present limitations that make them inappropriate for real-time fully automatic applications. This paper proposes a new approach for estimating the 3D pose of a traffic sign from a single monocular view. The proposed solution exploits the planarity of the signs, reducing their mapping to the image space to a planar homography. The estimation of the projective homography is achieved through the detection and matching of keypoint features. A thorough evaluation of 15 detectors and 19 detector/descriptor pairs is conducted to extract the best combination for each class of traffic signs. Obtained results, on challenging data generated from the German Traffic Sign Recognition benchmark (GTSRB), show the capacity of the proposed solution in dealing with the majority of traffic sign categories. We also propose amelioration to the evaluation metrics used in previous works namely the repeatability and matching scores.