基于可调场景语义标注范围的家庭室内语义地图构建
Semantic map construction for indoor home environment based on adjustable scene semantic annotation scope
张淑珍 1何镇 1查富生 2侯致远 1马玉祥1
作者信息
- 1. 兰州理工大学 机电工程学院,兰州 730000
- 2. 哈尔滨工业大学 机器人技术与系统国家重点实验室,哈尔滨 150001
- 折叠
摘要
针对家庭室内环境语义地图建图速度较慢和在门口场景语义标注易出现错误等问题,提出一种基于可调场景语义标注范围的家庭室内语义地图构建方法.首先根据 YOLOv5s 识别的物体大小赋予相应的场景置信度,基于该场景置信度设置阈值使得语义标注范围限制在机器人当前所在区域,确保场景切换时语义标注范围不会立即改变.然后基于人工势场虚拟力"引力斥力"原理,实现语义标注范围的扩大或缩小.最后结合阈值和动态语义标注范围,避免在门口场景中出现语义标注错误.实验结果表明:与Places205-VGG16神经网络建立家庭室内语义地图相比,所提方法平均效率和平均精准率分别提升了 11.0%和 7.8%,在家庭室内环境中具有一定的优越性.
Abstract
Aimed at the problems of slower construction speed of semantic map for home indoor environment and easy error in doorway scene semantic annotation,a semantic map construction method for indoor home environment based on adjustable scene semantic annotation scope is proposed.Firstly,the corresponding scene confidence is assigned according to the size of the object identified by YOLOv5s,and the threshold is set based on the scene confidence to limit the semantic annotation scope to the region where the robot is currently located,which ensures that the semantic annotation scope will not be changed immediately when the scene is switched.Then,based on the principle of"gravitational repulsion"of artificial potential field virtual force,the semantic annotation scope can be enlarged or narrowed.Finally,the threshold and dynamic semantic annotation scope are combined to avoid semantic annotation errors in the doorway scenarios.The experimental results show that compared with Places205-VGG16 neural network to build home indoor semantic maps,the proposed method improves the average efficiency and average accuracy by 11.0%and 7.8%respectively,which has certain superiority in home indoor environment.
关键词
家庭室内环境/语义地图/场景识别模型/场景置信度/变语义标注范围Key words
home indoor environment/semantic map/scene recognition model/scene confidence/variable semantic annotation scope引用本文复制引用
基金项目
国家重点研发计划"智能机器人"重点专项(2020YFB13134)
出版年
2024
NETL
NSTL
维普
万方数据