A Novel Small Vehicle Detection Method Based on UAV Using Scale Adaptive Gradient Adjustment

摘要

Vehicle detection based on UAV video is a typical small object detection task. In recent years, multi-scale prediction framework has become one of key steps for small object detection. However, the performances of existing methods are still not satisfactory for small object detection. In this paper, inspired by that the scale of object has an impact on gradient descent in the deep learning process, we choose the intersection over union (IOU) as the evaluation metric to analyze the relationship between scale of objects and gradient. We have shown that the gradient adjustment methods should satisfy some rules and thus we propose a new gradient adjustment formula based on our analysis. In addition, we built a mixed small vehicle dataset based on UAV videos for better evaluation of small vehicle detection. In the comparison with existing methods, our proposed method has achieved better results. The performance of our method reveals the potential of scale adaptive gradient descent method.