基于级联式Snappy-CenterNet 的锥套目标检测算法

针对空中加油因场景光照变化、环境遮挡等情况造成的锥套目标识别精度低、实时性差的问题，提出一种基于级联式Snappy-CenterNet深度网络的锥套目标检测算法。在CenterNet网络的基础上，以HourglassNet为主干网络，改进其bottleneck结构并引入中心池化的方法，对整体的网络结构进行优化，通过级联式的网络提升整体检测精度。实验结果表明：该算法可实现在多种复杂场景下对锥套目标的可靠检测，检测结果的精确率与召回率均可达99%，位置精度与区域精度分别可达99%与96%，更新率可达33.68 Hz，满足空中加油近距视觉导航阶段对于锥套识别的指标要求。

Drogue Target Detection Algorithm Based on Cascaded Snappy-CenterNet

Aiming at the problems of low accuracy and poor real-time performance of drogue target recognition in aerial refueling caused by scene illumination changes and environmental occlusion, a drogue target detection algorithm based on cascaded Snappy-CenterNet deep network is proposed. On the basis of CenterNet network, HourglassNet is used as the backbone network, the bottleneck structure is improved and the central pooling method is introduced to optimize the overall network structure, and the overall detection accuracy is improved through the cascaded network. The experimental results show that the proposed algorithm can reliably detect drogue targets in a variety of complex scenes, and the precision and recall of the detection results can reach 99%, the position accuracy and region accuracy can reach 99% and 96%, respectively, and the update rate can reach 33.68 Hz, which meets the requirements of aerial refueling near vision navigation for drogue recognition.