摘要:近年来，无人机技术的快速发展使得无人机地面目标检测技术成为计算机视觉领域的重要研究方向，无人机在军事侦察、交通管制等场景中具有普遍的应用价值.针对无人机场景下目标分辨率低、尺度变化大、相机快速运动、目标遮挡和光照变化等问题，提出一种基于残差网络的航拍目标检测算法.在SSD（single shot multibox detector）目标检测算法的基础上，用表征能力更强的残差网络进行基准网络的替换，用残差学习降低网络训练难度，提高目标检测精度；引入跳跃连接机制降低提取特征的冗余度，解决层数增加出现的性能退化问题.同时，针对SSD目标检测算法存在的目标重复检测和小样本漏检问题，提出一种基于特征融合的航拍目标检测算法.算法引入不同分类层的特征融合机制，把网络结构中低层视觉特征与高层语义特征有机地结合在一起.实验结果表明，算法在检测准确性和实时性方面均具有较好的表现.



Abstract:In recent years, the rapid development of UAV (Unmanned Aerial Vehicle) technology makes UAV ground target detection technology become an important research direction in the field of computer vision. UAV has a wide range of applications in military investigation, traffic control, and other scenarios. Nevertheless, the UAV images have many problems such as low target resolution, scale changes, environmental changes, multi-target interference, and complex background environment. Aiming at the above difficulties, derived from the original SSD target detection algorithm, this study uses a residual network with better characterization ability to replace the basic network and a residual learning to reduce the network training difficulty and improve the target detection accuracy. By introducing a hopping connection mechanism, the redundancy of the extracted features is reduced, and the problem of performance degradation after the increase of the number of layers is solved. The effectiveness of the algorithm is verified through experimental comparison. Aiming at the problem of target repeated detection and small sample missing detection of the original SSD target detection algorithm, this study proposes an aerial target detection algorithm based on feature information fusion. By integrating information with different feature layers, this algorithm effectively makes up for the difference between low-level visual features and high-level semantic features in neural networks. Results show that the algorithm has sound performance in both detection accuracy and real-time performance.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5695