Research on Motion Trend Enhanced 2D Detection on Drones

Abstract

Inspired by the human visual system, we proposed a motion information-based enhancement mechanism for drone detection, named Collaborative Filtering Mechanism (CFM). CFM enhances small object features through GAN-based image translation which is based on a Cycle Generative Adversarial Network (CycleGAN), and filters out unrelated features during the feature extraction cascade of YOLO-V5s, thus improving the performance of object detection. In the experiments, we verified the performance improvement brought by the proposed CFM module on the VisDrone dataset.