Leveraging cross-view geo-localization with ensemble learning and temporal awareness

Abstract

The Global Navigation Satellite System (GNSS) is unreliable in some situations. To mend the poor GNSS signal, an autonomous vehicle can self-localize by matching a ground image against a database of geotagged aerial images. However, this approach has challenges because of the dramatic differences in the viewpoint between aerial and ground views, harsh weather and lighting conditions, and the lack of orientation information in training and deployment environments. In this paper, it is shown that previous models in this area are complementary, not competitive, and that each model solves a different aspect of the problem. There was a need for a holistic approach. An ensemble model is proposed to aggregate the predictions of multiple independently trained state-of-the-art models. Previous state-of-the-art (SOTA) temporal-aware models used heavy-weight network to fuse the temporal information into the query process. The effect of making the query process temporal-aware is explored and exploited by an efficient meta block: naive history. But none of the existing benchmark datasets was suitable for extensive temporal awareness experiments, a new derivative dataset based on the BDD100K dataset is generated. The proposed ensemble model achieves a recall accuracy R@1 (Recall@1: the top most prediction) of 97.74% on the CVUSA dataset and 91.43% on the CVACT dataset (surpassing the current SOTA). The temporal awareness algorithm converges to R@1 of 100% by looking at a few steps back in the trip history.