Map-Matching-Based Cascade Landmark Detection and Vehicle Localization

For vehicle localization in highway situations, this paper proposes a map-matching based road facility detection and vehicle localization method with a camera, a low-cost global navigation satellite systems combined with an inertial navigation system (GNSS/INS), and a digital map. The proposed method adopts the cascade structure, which detects a road facility in each stage and gradually reduces the localization uncertainty based on the detection results. The proposed method consists of two stages. In the first stage, lane endpoints are detected and one camera position hypothesis for each lane is generated based on the detected lane endpoints. The localization uncertainty is reduced from a few meters to tens of cm. In the second stage, road sign regions of interest (ROI) are generated by projecting road signs of the map to an image based on each camera position hypothesis. Road signs are detected effectively within these ROIs and the best camera position hypothesis is selected based on the detection results. The proposed method significantly reduces the processing time and largely improves the road sign detection performance. Its processing time to detect a road sign in an image, whose resolution is 1280 × 1024 , is 13ms on average without the help of any parallel processing H/W such as a graphic processing unit. Its detection performance has 100% recall and 100% precision, and this result is superior to that of the deep neural network detector whose name is You Only Look Once version 3 (YOLOv3). The localization precision of the proposed method is 20cm in average in highway situations.