Lifelong Direct Error-Driven Learning for UAV Altitude Estimation in Different Weather Conditions

While deep neural networks achieve remarkable visual perception capabilities for UAV position and orientation estimation, their resilience to different weather conditions still needs improvement. These models often suffer from catastrophic forgetting when adapted to new environments, losing previously acquired knowledge. Lifelong learning methods aim to balance learning flexibility and memory stability. In this paper, we present an image-based approach to estimate the relative altitude of a UAV using 2D images under varying weather conditions, including sunny, sunset, and foggy scenarios. Our experiments demonstrate significant performance degradation when the model is trained sequentially on different weather datasets, especially when new images differ substantially from those in the initial training dataset. However, testing Elastic Weight Consolidation (EWC) and Direct Error-Driven Learning (EDL) separately showed that each method helps maintain stability and performance across various weather conditions. Our results show the feasibility and effectiveness of these methods in diverse environmental conditions.