Multi-view traffic sign localization with high absolute accuracy in real-time at the edge

In this paper, we present a system to localize traffic signs with high accuracy in real-time at the edge using 3D reconstruction of an environment. We use Structure-from-Motion (SfM) based 3D reconstruction with only a small number of tracked feature points in video frames. The 3D model obtained from SfM alone has an arbitrary scale and orientation, which is not suitable for localizing traffic signs absolutely. Therefore, we use the GNSS locations of the images to scale and orient the 3D model with the real world. Next, we compute the latitude-longitude location of a traffic sign using the mapping between the traffic sign bounding box and corresponding 3D points in the 3D model. We also present the design of other relevant system components such as vehicle location, timing, data synchronization, and sensor calibration, which are required to achieve high accuracy localization. We evaluated our system in city, suburb, and highway scenes on meticulously annotated ground truth datasets. According to evaluation, our system achieves the median localization accuracy of 2.76 meters with high localization success rate and runs on average 15× faster than a comparable baseline system.