Object Detection in UAV Aerial Images Based on Improved YOLOv7-tiny

作者：Ziyang Wang, Zhiwen Liu, Genqi Xu, Siyuan Cheng · 年份：2023 · DOI：10.1109/cvidl58838.2023.10166362 · 被引用次数：9 · 研究领域：Advanced Neural Network Applications、Robotics and Sensor-Based Localization、Advanced Image and Video Retrieval Techniques

With the wide application of unmanned aerial vehicle (UAV), the problem of object detection in UAV aerial image has attracted more and more attention. In order to detect objects in UAV aerial images more accurately, we made a series of improvements on the basis of YOLOv7-tiny to make it more suitable for detecting objects in UA V aerial images. Specifically, we added Global Attention Mechanism (GAM) to YOLOv7-tiny's neck network to better extract the features of the objects. We also introduced the Bidirectional Feature Pyramid Network (BiFPN) into the Neck network to enhance feature fusion capability. In order to detect small objects more accurately, we added a small object detection head in the YOLOv7-tiny Head section. Contextual Transformer module was introduced to make the network pay more attention to context information. To make the predicted box match the ground truth box more closely, we applied the SIoU loss function to YOLOv7-tiny. The experimental results showed that, mAP@0.5 is 38.1 % and mAP@0.5:0.95 is 21.3% by improved YOLOv7-tiny algorithm on VisDrone dataset. Compared with YOLOv7-tiny, mAP@0.5 is improved by 2.8% and mAP@0.5:0.95 is improved by 3%, which improved the accuracy of object detection in UAV aerial images.