电子测量与仪器学报

刘紫燕,杨模,袁浩,梁静,梁水波,孙昊堃.结合拆分注意力机制和下一次预期观察的视觉导航[J].电子测量与仪器学报,2023,37(1):96-105

结合拆分注意力机制和下一次预期观察的视觉导航

Visual navigation combining split attention mechanism and next expected observation

DOI：10.13382/j.issn.1000-7105.2023.01.011

英文关键词:visual navigation deep reinforcement learning split attention mechanism next expected observations (NEO)

基金项目:贵州省科学技术基金(黔科合基础[2016]1054)、贵州省联合资金( 黔科合 LH 字[2017]7226 号)、贵州大学 2017 年度学术新苗培养及创新探索专项( 黔科合平台人才[2017]5788)、贵州省科技计划项目(黔科合 SY 字[2011]3111)资助

作者	单位
刘紫燕	1. 贵州大学大数据与信息工程学院，2. 贵州大学大数据国家重点实验室
杨模	1. 贵州大学大数据与信息工程学院
袁浩	1. 贵州大学大数据与信息工程学院
梁静	1. 贵州大学大数据与信息工程学院
梁水波	1. 贵州大学大数据与信息工程学院
孙昊堃	1. 贵州大学大数据与信息工程学院

Author	Institution
Liu Ziyan	1. College of Big Data and Information Engineering, Guizhou University,2. State Key Laboratory of Public Big Data, Guizhou University
Yang Mo	1. College of Big Data and Information Engineering, Guizhou University
Yuan Hao	1. College of Big Data and Information Engineering, Guizhou University
Liang Jing	1. College of Big Data and Information Engineering, Guizhou University
Liang Shuibo	1. College of Big Data and Information Engineering, Guizhou University
Sun Haokun	1. College of Big Data and Information Engineering, Guizhou University

摘要点击次数: 651

全文下载次数: 675

中文摘要:

针对深度强化学习视觉导航算法因导航场景变化而导致导航精度下降,影像匹配的实时性和可靠性降低的问题,提出一种融合拆分注意力机制和下一次预期观测(NEO)的视觉导航模型。首先使用 ResNest50 骨干网络提取当前状态和目标状态的特征以降低网络冗余,利用跨阶段部分连接 CSP 强化捕获浅层目标特征信息以增强模型的学习能力。然后提出改进的损失函数,使得推理网络更加接近于真实后验,从而智能体能在当前环境下做出最佳决策,进一步提升不同场景下模型的导航精度。在 AVD 数据集和 AI2-THOR 场景进行训练及测试,实验结果表明,本文算法导航精度高达 96. 8%,平均 SR 提升约 3%,平均 SPL 提升约 6%,可以满足导航精度和实时匹配的要求。

英文摘要:

A visual navigation model incorporating split attention mechanism and next expected observation ( NEO) is proposed to address the problem that deep reinforcement learning visual navigation algorithm degrades navigation accuracy, real-time and reliability of image matching due to navigation scene changes. The features of current and target states are first extracted using the ResNest50 backbone network to reduce network redundancy. The shallow target feature information is captured intensively using a cross-stagepartial-connections CSP to enhance the learning ability of the model. Then an improved loss function is proposed to make the inference network closer to the true posterior so that the agent can make the best decision in the current environment and further improve the navigation accuracy of the model in different scenarios. The training and testing are conducted on AVD dataset and AI2-THOR scenes, and the experimental results show that the navigation accuracy of the algorithm in this paper is as high as 96. 8%, with an average SR improvement of about 3% and an average SPL improvement of about 6%, which meets the requirements of navigation accuracy and realtime matching.

查看全文查看/发表评论下载PDF阅读器