An Experimental Study of Vehicle Detection on Aerial Imagery Using Deep Learning-Based Detection Approaches

Abstract Deep convolutional neural network technology is widely used to deal with general object detection in computer vision, and it achieved remarkable progress. Unmanned aerial vehicles provide large numbers of aerial imagery that significantly facilitate several applications including traffic monitoring, surveillance, tracking, rescue, and safe military tasks. This paper presents an experimental study to evaluate the performances of several state-of-the-art deep learning-based detection approaches on vehicle detection from aerial imagery. The pre-trained models, including Faster R-CNN, R-FCN, and SSD, are adopted from the TensorFlow model zoo, and the VEDAI dataset is used as the benchmark. The results show that Faster R-CNN combined with Resnet101 backbone achieved the highest mAP, which is 39.73% on the COCO metric. This experimental study expects to be a guideline to choose suitable approaches for particular applications.