Optimizing Object Detection in Remote Towers: A Self-Improving Approach through Adaptive Annotation and Model Re-Training

With the widespread adoption of multi-airport remote tower systems, the demand for panoramic object detection in complex environments has increased significantly. However, traditional object detection techniques heavily rely on manually collected and annotated datasets for model training, which suffer from limited generalization capability and inefficient annotation processes. These limitations hinder the accuracy of detection models across diverse scenarios. This paper proposes an adaptive data annotation and training method aimed at enhancing the robustness and automation of detection models in dynamic environments. The proposed approach automatically identifies and stores data associated with poor detection performance during real-world operations. It then corrects and retrains the model iteratively, allowing for continuous optimization of model performance. Experimental results demonstrate that, compared to conventional methods, this approach significantly improves the generalization and detection accuracy of models across varying weather conditions and scenes, while reducing reliance on human intervention. This presents significant practical value for engineering applications.