Airport Scene Aircraft Detection Method Based on YOLO v3

Authors:
  • 郭进郭进祥GuoJinxiang
    郭进祥GuoJinxiang
  • 刘立刘立波LiuLibo
    刘立波LiuLibo
  • 徐峰徐峰XuFeng
    徐峰XuFeng
  • 郑斌郑斌ZhengBin
    郑斌ZhengBin

Abstract
The difficulty to detect small targets or occlusion aircrafts poses a great challenge to the accuracy and real-time of aircraft detection. In this paper, YOLO v3 algorithm with high real-time performance is applied to the field of aircraft detection in airport scene, and two improvements are proposed: replacing the convolution layer in backbone network with void convolution, maintaining high resolution and large field of receptivity and improving the accuracy of small target detection; optimizing the NMS algorithm by linear attenuation confidence score to improve the detection accuracy of occlusion aircrafts. The results show that the improved YOLO v3 can well detect small targets and occlusion aircraft, and the detection accuracy is improved from 72.3% to 83.7% as the real-time performance is ensured.