Towards Safer Roads: A Data-Driven Ground-Aerial Fusion Approach for Enhanced Ego-Localization

Kurzfassung

Accurate localization is crucial for autonomous systems to ensure safe and effective operation on modern roadways. While consumer-grade GPS sensors are widely used, they often fall short due to limitations imposed by satellite geometry, atmospheric disturbances, and multipath interference. These factors can result in errors of several meters, a discrepancy that is unacceptable for the centimeter-level accuracy required for autonomous driving. Inaccurate positioning not only affects navigation, but also increases the risk of collisions, thus undermining road safety. To overcome these challenges, autonomous vehicles are now integrating a variety of sensors such as inertial measurement units, LiDAR, radar, and computer vision to provide a more reliable picture of the environment. These systems combine data streams to improve overall accuracy and resilience to individual sensor failures. Filtering algorithms, including Kalman and particle filters, have been instrumental in refining localization estimates. Computer vision techniques such as visual odometry further mitigate scenarios. More recently, deep learning approaches have emerged as powerful tools capable of learning complex mappings between sensor inputs and precise location outputs. These advances not only enhance autonomous navigation, but also contribute to road safety by reducing the risk of accidents and promoting smoother interactions among all road users. In this work, we develop a robust solution for ego-localization. We propose to create a carefully curated dataset that overcomes current shortcomings by providing centimeter-level GPS annotations, diverse scenarios, and rich temporal dynamics. In addition, we design an innovative architecture that fuses ground-view images with rough GPS localization data to accurately predict true positions.