基于对比学习的轻量化弱监督SLAM闭环回路检测  

Lightweight Weakly Supervised SLAM Loop Closure Detection Based on Contrastive Learning

作　　者：王传云[1] 娄渊伟 刘晓娜 王静静 高骞 WANG Chuanyun;LOU Yuanwei;LIU Xiaona;WANG Jingjing;GAO Qian(Shenyang Aerospace University,Shenyang 110136,China;Yantai Science and Technology Innovation Promotion Center,Yantai 264003,China;China Academic of Electronics and Information Technology,Beijing 100041,China)

机构地区：[1]沈阳航空航天大学,沈阳110136 [2]烟台市科技创新促进中心,烟台264003 [3]中国电子科技集团公司电子科学研究院,北京100041

出　　处：《空间控制技术与应用》2023年第5期47-54,共8页Aerospace Control and Application

基　　金：国家自然科学基金资助项目(61703287);辽宁省教育厅科学研究项目(LJKZ0218和LJKMZ20220556);沈阳市中青年科技创新人才项目(RC210401);沈阳航空航天大学引进人才科研启动基金项目(22YB03)。

摘　　要：为了构建在视觉定位过程中时隔更加久远的约束,便于搭载视觉定位设备的飞行器能够建立全局一致的轨迹估计,提出一个基于对比学习的轻量化弱监督即时定位与地图重建技术(simultaneous localization and mapping,SLAM)闭环回路检测算法.通过对图像建立一致的全局描述符,利用相似度度量判断轨迹中的闭环回路,建立长时序约束.考虑到在资源受限平台上的应用,利用EfficientNet(efficient neural network)实现更高效的特征提取,并结合需求压缩与激励(need squeeze and excitation,NSE)注意力模块提升数据降维过程中对有效数据的筛选,通过完整的局部聚合描述符矢量(vector of locally aggregated descriptors,VLAD)层整合全局描述符,使网络模型最终能够在光照条件、视角和季节等环境变化中仍具备高效的识别能力.实验结果表明,在保持与基线模型TOP-5召回率指标相差2%的基础上,所提出的方法能够有效缩减57%模型体积、35%模型训练时间并提升48%执行效率,有利于部署在小型无人机等资源受限的嵌入式平台上.In order to construct a longer-term constraint in the process of visual positioning,and facilitate the estab-lishment of globally consistent trajectory estimation for aircraft equipped with visual positioning equipment,a light-weight weakly supervised simultaneous localization and mapping(SLAM)loop closure detection algorithm is pro-posed based on contrastive learning.By establishing consistent global descriptors for images and using similarity measures to judge loops in trajectories,long-term timing constraints are established.Considering the application on resource constrained platforms,an EfficientNet(efficient neural network)is used to achieve more efficient feature extraction,and the NSE(need squeeze and excitation)attention module is combined to improve the screening of effective data in the process of data dimensionality reduction.Global descriptor is integrated through a complete VLAD(vector of locally aggregated descriptors)layer.The network model can still have efficient recognition ability in environmental changes such as lighting conditions,viewing angles,and seasons.The experimental results show that,while maintaining a 2%difference in the TOP-5 recall index compared to the baseline model,the proposed method can effectively reduce model volume by 57%,training time by 35%,and improve execution efficiency by 48%,which is beneficial for deployment on resource constrained embedded platforms such as small drones.

关 键 词：闭环回路 轻量化 弱监督 对比学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]