Object Detection in Drone Video with Temporal Attention Gated Recurrent Unit Based on Transformer

Abstract

Unmanned aerial vehicle (UAV) based object detection plays a pivotal role in civil and military fields. Unfortunately, the problem is more challenging than general visual object detection due to the significant appearance deterioration in images captured by drones. Considering that video contains more abundant visual features and motion information, a better idea for UAV based image object detection is to enhance target appearance in reference frame by aggregating the features in neighboring frames. However, simple feature aggregation methods will frequently introduce the interference of background into targets. To solve this problem, we proposed a more effective module, termed Temporal Attention Gated Recurrent Unit (TA-GRU), to extract effective temporal information based on recurrent neural networks and transformers. TA-GRU works as an add-on module to bring existing static object detectors to high performance video object detectors, with negligible extra computational cost. To validate the efficacy of our module, we selected YOLOv7 as baseline and carried out comprehensive experiments on the VisDrone2019-VID dataset. Our TA-GRU empowered YOLOv7 to not only boost the detection accuracy by 5.86% in the mean average precision (mAP) on the challenging VisDrone dataset, but also to reach a running speed of 24 frames per second (fps).