点击次数：
所属单位：电子信息工程学院
发表刊物：Int. Conf. Syst. Informatics, ICSAI
摘要：Although Faster R-CNN has excellent performance in object detection, it still has some difficulties in detecting small targets and slightly overlapped targets in UAV (Unmanned Aerial Vehicle) images. Based on Faster R-CNN, this paper uses ResNet101 as a feature extractor. We increase the number of anchors from 9 to 15 in RPN so that the small targets can match more anchors and get sufficient training. Due to the increasement of anchors, this paper introduces a 1\times1 convolution layer to integrate features and reduce the feature map channels. We also apply RoIAlign to avoid the misalignment caused by RoIPool. The improved model effectively increases the detection rate of small targets and slightly overlapped targets so that it can be applied to human detection under UAV. The improved model can detect small targets with a size of about 30\times80 pixels on aerial images with resolution of 3840\times2160 pixels. Compared with Faster R-CNN, the improved model increases AP (Average Precision) from 74.31% to 79.77% on the WILDTRACK dataset. © 2018 IEEE.
是否译文：否
发表时间：2019-01-02
合写作者：Zhu, Hanshan,Qi, Yayun,Shi, Haochen,Zhou, Huiyu
通讯作者：黎宁