Improving Vehicle Localization with Lane Marking Detection Based on Visual Perception and Geographic Information

With recent advance of deep learning techniques, the development of intelligent vehicles is constantly moving towards fully autonomous driving. Many essential functions of advanced driver assistance system (ADAS) have been well investigated. In this paper, we address the problem of accurate vehicle localization with visual perception and geographic information. The proposed method combines drivable area detection, lane line detection and image-map matching to achieve lane-level localization accuracy. We present a confidence estimation approach to generate virtual lane lines for lane marking detection. In addition, a framework for map database update is implemented with vehicle localization and traffic light detection. The proposed networks are pre-trained using CULane and fine-tuned on our dataset. In the experiments, the localization accuracy is evaluated using MAE and RMSE. The results have demonstrated the improvement over the GPS- based methods.