摘要
针对家庭环境中目标检测遇到的常见问题,如物体特征较少、类别复杂多样、位置重叠密集等,当前通用的基于深度学习的二阶段网络Faster R-CNN或者一阶段网络YOLO系列都无法学习到足够的特征来区分这些复杂物品,性能表现不佳。将基于图卷积神经网络GCN的算法空间感知图关系网络引入目标检测系统中,利用CNN网络提取图像特征后,在候选框ROI结构后嵌入图网络模块。利用邻接矩阵和高斯核同时建模目标的语义和空间信息,学习目标物品在家庭环境中的空间因果对应关系,通过图结构推理增强后续识别检测效果。实验结果表明,将图网络方法应用到服务机器人平台是可行的,在保持检测速度的前提下检测精度更高,且具有较好的鲁棒性和泛化能力。
In view of the common problems encountered in target detection in the home environment, such as fewer object features, complex and diverse categories, dense overlap, etc. the current general deep learning-based two-stage network Faster R-CNN or the one-stage network YOLO series cannot learn with enough features to distinguish these complex items, the performance is poor. The algorithm space-aware graph relation network based on GCN is introduced into the target detection system. The CNN network is used to extract image features, and the graph network module is embedded in the ROI. The adjacency matrix and the Gaussian kernel are used to simultaneously model the semantic and spatial information of the target, learn the spatial causal relationship of the target item in the home environment, and enhance the subsequent recognition and detection effect through graph structure reasoning. The experimental results show that it is feasible to apply the graph network method to the service robot platform, the detection accuracy is higher while the detection speed is maintained, and has better robustness and generalization ability.
作者
陈鑫
张奇志
周亚丽
Chen Xin;Zhang Qizhi;Zhou Yali(School of Automation,Beijing Information Science&Technology University,Beijing 100192)
出处
《中国仪器仪表》
2021年第9期49-55,共7页
China Instrumentation
基金
国家自然科学基金资助项目(11672044)。
关键词
服务机器人
目标检测
图神经网络
空间语义推理
Service robot
Object detection
Graph neural network
Spatial semantic reasoning
