Incorporating Onboard Visual Sensing and Geographic Information for Lane-Level Vehicle Localization

With the recent advances in deep learning and artificial intelligence, vehicular technologies are progressing rapidly toward fully autonomous driving. Critical components of advanced driver assistance systems (ADASs) have been extensively investigated in the past decade. In this article, we address the challenge of precise vehicle localization using visual perception and geographic information. The proposed method incorporates drivable area detection, lane line detection, and image-map matching to achieve lane-level localization accuracy. It takes advantage of digital maps and images from the onboard cameras to simultaneously optimize vehicle localization and lane detection. A confidence estimation is proposed to generate virtual lane lines for lane marking identification. Our networks are pretrained on the Cityscapes, CULane, and TuSimple datasets, and finetuned with the newly proposed TRoad and TLane. In the experiments, the performance evaluated using mean absolute error (MAE) and root mean square error (RMSE) have demonstrated a significant improvement over the conventional global positioning system (GPS)/global navigation satellite system (GNSS)-based vehicle localization techniques. Source code and datasets are available at https://github.com/Wally0924/Enhancing-Vehicle-Localization-Using-Lane-Detection-and-Geographic-Information