Anchored versus Anchorless Detector for Car Detection in Aerial Imagery

Abstract

With the increase in the traffic on roadways, traffic monitoring is the major need we have at this moment. Using UAVs for traffic monitoring has numerous advantages such as broader field of view, higher mobility, no effect on detected traffic, etc., however, variation in camera orientation, UAV height, cluttered background imposes challenges to this aerial object detection. To provide a UAV-based traffic monitoring solution, we have proposed a car detection system for UAV images using deep learning approaches. We compared the performance of the anchorless Fully Convolutional One Stage (FCOS) object detection algorithm with the popular YOLOv3 algorithm. The performance analysis of these models based on mean Average Precision (mAP) indicates that FCOS yields better results over YOLOv3, whereas in terms of computation speed YOLOv3 performed better.