适用于非合作目标捕获的轻量级位姿估计网络  

Lightweight Pose Estimation Network for Non-Cooperative Target Acquisition

作　　者：蒋明 陈雨[1] 周青华 袁媛[1] 何世琼 JIANG Ming;CHEN Yu;ZHOU Qinghua;YUAN Yuan;HE Shiqiong(School of Electronic Information,Sichuan University,Chengdu 610065,China;School of Aeronautics and Astronautics,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,成都610065 [2]四川大学空天科学与工程学院,成都610065

出　　处：《计算机工程》2022年第6期235-242,共8页Computer Engineering

基　　金：国家自然科学基金面上项目(51875373);四川省科技计划项目(2019YJ0093)。

摘　　要：空间非合作目标的增多导致太空安全受到严重威胁,对非合作目标进行捕获回收具有维护空间安全、节约资源等现实意义。非合作目标捕获回收需要进行位姿估计,而目前在硬件资源有限的航天器平台上,现有的大多数非合作目标位姿估计算法无法同时满足及时性和准确性的要求。设计一种超轻量级目标检测网络YOLOGhostECA,利用GhostBottleneck网络减少特征图冗余,并使用高效注意力机制提取核心特征图,以降低模型参数,在提升运算速度的同时保证精度水平几乎不下降。根据YOLO-GhostECA网络的检测结果粗略估计姿态,以协助机械臂更加合理地执行智能捕获任务,解决2D识别算法无法检测出物体姿态的问题。在7自由度冗余机械臂上开展的空间非合作目标捕获地面模拟的实验结果表明,与YOLOv5s网络相比,该网络模型大小减小了80.4%,运算复杂度降低了78.9%,而精度基本保持不变,可准确快速地对非合作目标进行位姿估计,能够引导机器人成功捕获非合作物体。The increase in non-cooperative targets in space has led to a serious threat to space security.Therefore,the capture and recovery of non-cooperative targets have practical significance in maintaining space security and resourcesaving.Non-cooperative target acquisition and recovery require pose estimation.On a spacecraft platform with limited hardware resources,most existing non-cooperative target pose estimation algorithms cannot simultaneously meet the requirements of timeliness and accuracy.An ultra-lightweight target detection network YOLO-GhostECA is proposed.The Ghostbottleneck network reduces the redundancy of the feature map,and an efficient attention mechanism extracted the core feature map to reduce the model parameters and ensure that the accuracy level was not compromised while improving the operation speed.Using the detection results of the YOLO-GhostECA network to estimate the attitude to assist the manipulator in a sensibly intelligent acquisition task and solve the problem that the 2D recognition algorithm cannot detect the object attitude.The experimental results of the ground simulation of space non-cooperative target acquisition on a 7-DOF redundant manipulator show that compared with the algorithm based on the YOLOv5s network,the model size was reduced by 80.4%,the computational complexity was reduced by 78.9%,and the accuracy remained constant.It accurately and quickly estimates the pose of non-cooperative targets and guidesthe robot to successfully capture non-cooperative objects.

关 键 词：非合作目标 神经网络 目标检测 智能捕获 轻量级 位姿估计 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]