Crowdsourced Road Semantics Mapping Based on Pixel-Wise Confidence Level

Abstract

High-definition map has become a vital cornerstone in the navigation of autonomous vehicles in complex traffic scenarios. Thus, the construction of high-definition maps has become crucial. Traditional methods relying on expensive mapping vehicles equipped with high-end sensor equipment are not suitable for mass map construction because of the limitation imposed by its high cost. Hence, this paper proposes a new method to create a high-definition road semantics map using multi-vehicle sensor data. The proposed method implements crowdsourced point-based visual SLAM to align and combine the local maps derived by multiple vehicles. This allows users to modify the extraction process by using a more sophisticated neural network, thus achieving a more accurate detection result when compared with traditional binarization method. The resulting map consists of road marking points suitable for autonomous vehicle navigation and path-planning tasks. Finally, the method is evaluated on the real-world KAIST urban dataset and Shougang dataset to demonstrate the level of detail and accuracy of the proposed map with 0.369 m in mapping errors in ideal condition.