Performance Analysis of Vehicle Detection Algorithm in Aerial Traffic Videos

This paper aims at the problem of vehicle detection in UAV aerial videos and tries to investigate the performance of different kinds of object detection algorithms under such a condition. We chose ViBe, HOG+SVM, Faster R-CNN and YOLOv3 as the typical detection algorithms. After applying them on UAV aerial videos with different height and traffic scenes and testing their precision and recall, we make a theoretical analysis of their performance. Our experiments showed that Faster R-CNN is the best method. Meanwhile, ViBe, with the advantages of less calculation and relatively low hardware requirements, is a good choice for a simple scene where the vehicles keep moving all the time.