Abstract: The requirements by autonomous driving to HD maps, including high-precision, high-quality, and vast-coverage, which is less likely to be fully addressed with current technology. This paper proposes a method to convert low-precision observations into confidence of high-precision map elements, for they are different in spatial but very similar at feature level. It improves the strategy of HD map reliability without shorten its update cycle. The process starts with narrowing down matching scope and measuring matching rate, twice filtering using rectangular neighbor and a measuring model are employed to figure out the target from HD map. Followed by feature level analysis, status assessment and confidence score calculate, confidence of HD map elements get updated afterwards. Applying this method to our experiments with HD map includes 17 provinces, confidence of traffic sign gets updated via dash cam images. Moreover, two ways that could trigger out confidence update are verified, i.e. new observation and new map inputs. The results show redundant observations and confidence level are in positive correlation. When observation redundancy reaches 15, more than 80% of the target elements can maintain high confidence for 60 days, and the accuracy rate exceeds 96%. This study provides a solution to the conflict between high-precision map freshness and reliability, which can further promote the deepening of strategies of high-precision map applications.

Key words: high definition map, map confidence, map update, map matching