How to Match Tracks of Visual Features for Automotive Long-Term SLAM

Accurate localization is a vital prerequisite for future assistance or autonomous driving functions in intelligent vehicles. To achieve the required localization accuracy and availability, long-term visual SLAM algorithms like LLama-SLAM are a promising option. In such algorithms visual feature tracks, i.e. landmark observations over several consecutive image frames, have to be matched to feature tracks recorded days, weeks or months earlier. This leads to a more challenging matching problem than in short-term visual localization and known descriptor matching methods cannot be applied directly. In this paper, we devise several approaches to compare and match feature tracks and evaluate their performance on a long-term data set. With the proposed descriptor combination and masking ("CoMa") method the best track matching performance is achieved with minor computational cost. This method creates a single combined descriptor for each feature track and furthermore increases the robustness by capturing the appearance variations of this track in a descriptor mask.