A Comparative Study of Small Object Detection Algorithms

Abstract

The scenario considered in this paper is detecting small targets using an unmanned aerial vehicle(UAV) equipped with a pan-tilt-zoom(PTZ) camera. We reviewed, compared and modified the state-of-the-art target detection algorithms. Different target detection algorithms have different advantages in terms of accuracy and time consumption. The contributions of this paper are that we obtained pictures using the UAV and created a dataset firstly, then a support vector machine(SVM) based classifier and a YOLO-based Neural Network have been designed to detect targets separately, and both algorithms are evaluated on the test dataset finally. In this paper, the advantages and disadvantages of traditional machine learning methods based on SVM and deep learning methods based on YOLO on target detection are analyzed through a series of experiments. Limited by the target area selection methods, machine learning methods based on SVM do not have a high detection accuracy. The deep learning method based on neural network focus on improving the accuracy of target detection, while slightly inferior in time consumption. In addition, we modified the YOLO network structure for the special scene of small objects detection in this paper. Experiments show that the modified network can reduce the time consumption and improve the accuracy of small target detection. In summary, different types of target detection methods should be selected and modified according to the practical application scenarios.