Automated vehicle detection and instance segmentation from high-resolution UAV imagery using YOLOv7 model

Abstract

Automatic vehicle detection from unmanned aerial vehicles (UAVs) is an important task in the remote sensing domain and plays a pivotal role in many applications such as traffic monitoring, parking lot management, search and rescue tasks. Inspired by the success of the deep learning paradigm in image processing applications, many object detection, and tracking approaches have been developed and successfully employed in UAV-based object detection studies. In this study, automatic vehicle detection and instance segmentation was conducted using YOLOv7, which is the latest version of the You Only Look Once (YOLO) model from high-resolution UAV data obtained from Gebze Technical University campus in Turkey. For this purpose, vehicle images were collected from the UAV data of the study area, and the vehicles in the images were manually annotated with the LabelMe annotation tool. With the created dataset, the YOLOv7 algorithm was trained and tested with a transfer learning approach on Google Colab's virtual machine. Experimental results revealed that the YOLOv7 model achieved the Precision, Recall, and mAP@0.50 values for the bounding boxes and masks of vehicles as 99.79%, 97.54%, and 99.46%, respectively.