基于PIR的办公建筑室内人员定位及动作识别模型
Model for Location and Action Recognition of Indoor Occupant in Office Buildings Based on PIR Sensor Array
张静思 1赵婷 2周翔2
作者信息
- 1. 美的集团中央研究院,上海 201702
- 2. 同济大学机械与能源工程学院,上海 201804
- 折叠
摘要
准确识别人行为,是办公场景智能化的基础.本研究通过收集场景内的被动红外(PIR)传感器阵列信号,分析人员在不同位置、不同动作强度组合下的数据特征.首先,基于多种机器学习算法建立人员位置和动作同时识别的模型,相较于KNN、RF、SVM、MLP单一模型,Stacking融合模型性能更优、稳定性更好.然后,分析了面向连续数据流时,多时间步长Stacking模型组合表现更优;注重整体指标时,建议选用短时间步长组合模型;注重识别精度时,建议使用长时间步长组合模型.模型识别准确率达0.99,人员位置识别准确率为0.87,人员动作强度识别准确率为0.89.最后,讨论了区域划分和仪器布置对模型的影响.采用大尺度划分方式,人员位置识别准确率提高;小尺度划分方式下,人员定位更精确.可以通过调整PIR传感器的空间布置,减少传感器台数至8台.
Abstract
The accurate identification of occupant behavior is the foundation of smart workspace.This study employed a passive infrared(PIR)sensor array to monitor occupant in the scene and analyzes the data characteristics of individuals at different positions and action intensity combinations.A recognition model was established utilizing multiple machine learning algorithms.Compared with single models such as KNN,RF,SVM,and MLP,the Stacking fusion model demonstrates superior performance and stability.When confronted with continuous data streams,the Stacking model with multiple time step combinations outperforms others.Considering overall indicators,short time step combination models are recommended;considering identification accuracy,long time step combination models with higher action intensity recognition accuracy are recommended.The model achieves an identification accuracy of 0.99,occupant location identification accuracy of 0.87,and occupant action intensity identification accuracy of 0.89.Then,the impact of region division and instrument layout on the model's performance was discussed.Using large-scale partitioning,the accuracy of occupant location recognition was improved;under small-scale partitioning,the location of occupant is more precise.The number of sensor stations can be reduced to eight by adjusting the spatial layout of PIR sensors.
关键词
建筑人行为/PIR传感器/动作识别/机器学习/人员位置Key words
occupant behavior/PIR sensor/occupant action recognition/machine learning/occupant position引用本文复制引用
基金项目
国家重点研发计划(2017YFC0702200)
出版年
2024
NETL
NSTL
万方数据