Research And Application Of Online Recognition Method Of Human Thermal Sensation Based On Video Understanding

HVAC systems account for a large proportion of building energy consumption,and effective reduction of HVAC system energy consumption is of great significance in the context of the national goal of "carbon neutrality and carbon peaking".Most of the current heating and cooling are operated at a set value for a fixed period of time,which is set by the air conditioning manager or the user of the indoor environment according to the subjective thermal sensation of the moment,which does not take into account the thermal comfort of everyone in the room and leads to unnecessary waste of resources.Therefore,it is necessary to conduct an in-depth discussion and research on the real-time thermal sensation of occupants as well as the comprehensive thermal comfort.Although researchers have proposed a variety of thermal sensation prediction methods,there are still certain problems and limitations in this area of research:(1)The traditional thermal sensation evaluation models based on environmental thermal parameters mainly reflect the average thermal sensation statistical indicators obtained based on statistics,which are difficult to meet the individual needs of occupants.Such models are more accurate in environments with smaller wind speeds and air conditioning equipment.Moreover,the models do not consider the interaction information between occupants and the environment,the dynamic characteristics of human thermal comfort and the human thermal adaptability and regional differences in human thermal sensation,resulting in the poor generality and effectiveness of the model.(2)Smart terminal-based thermal sensation detection can maintain basic synchronization with personnel real-time thermal sensation,but there are defects such as high cost,inconvenience in wearing,and limited application.(3)Non-intrusive detection methods such as infrared sensor-based technologies have low adaptability due to their limited monitoring range,low detection accuracy and not applicable to personnel stationary detection and other shortcomings.The development of indoor environment control technology that maximizes the thermal comfort needs of multiple people and energy efficiency of buildings is extremely challenging due to a series of challenges such as non-energy saving,non-optimization of energy saving and difficulty in guaranteeing control performance of building indoor thermal environment operation control.One of the main technical bottlenecks to solve the above problems is the fast,accurate and convenient identification of real-time thermal sensation of occupants.In this paper,we extract the behavioral features generated by the thermal comfort adjustment process of occupant in the surveillance video,and use the perception capability of deep learning to study the online recognition technology of personnel thermal sensation based on surveillance video,and integrate personalized thermal comfort state feedback into the control of HVAC systems.This study analyzes the thermal environment and personnel as a whole system,systematically analyzes the characteristics of the "thermal environment + multiple people" composite system,establishes the feasibility of optical cameras(monitoring video)for thermal sensory state inference,provides new theories and technologies for the development of a new generation of artificial intelligence-driven real-time thermal sensory recognition technology,and provides new ideas for controlling the temperature setting of HVAC systems and reducing energy consumption.The specific research contents include:(1)Analyze the correlation between behavior and thermal sensation to create labels for thermal adaptation behavior categories.The collection of human behavior data is the basis for building thermal adaptive behavior,and thermal adaptive behavior categories are the necessary input for thermal sensory detection applications based on behavior recognition.First,thermal adaptation behavior mining is performed to mine the correlations between daily behaviors and heat sensations from behavioral data to identify the heat sensations that trigger behaviors.Second,based on the statistical data of behaviors and thermal sensations counted,suitable probabilistic or statistical models are investigated to establish correlations between behaviors and thermal sensations and generate models to transform behavioral data into descriptions of thermal sensations.(2)Establishing thermal adaptive behavior dataset.In the published datasets,the number of samples is too small and the quality level is not uniform,so the heat-adapted behavior recognition algorithms are mostly based on hand-designed features,which have large limitations and poor robustness.In order to implement the thermal adaptive behavior recognition algorithm based on deep convolutional networks,this paper collects and processes a sufficient amount of video data to form a large dataset of thermal adaptive behavior to provide a data base for the training of the algorithm.(3)A thermal adaptive behavior recognition framework based on human pose estimation is established to extract the behavior of characters in videos as a sequence of spatio-temporal locations of nodes,reduce the adverse effects of irrelevant color information and confusing background information in videos,and thus improve the computational efficiency of character action recognition in real-time surveillance videos.We select suitable training and test sets and develop reasonable criteria to objectively evaluate the performance of the thermal adaptive behavior recognition model in terms of accuracy,efficiency,robustness and feasibility.(4)Establishing behavioral feature descriptors with high discriminative power and extracting features with strong descriptive and judgmental power from videos are the keys to achieve thermal comfort state recognition.In this paper,we use an adaptive graph convolutional neural network combined with Transformer to provide key nodal features that play a decisive role in character behavior recognition for online video-based thermal sensory analysis,thereby improving the performance of thermal sensory recognition mechanisms and thus the accuracy of distinguishing thermal comfort states.(5)Establishing a thermal adaptive behavior recognition network.Due to the complexity of the actual scene environment,it is usually difficult to obtain the complete skeleton sequence.In order to effectively obtain the feature descriptors of incomplete skeleton and improve the accuracy of character thermal comfort state recognition,this paper proposes a multi-stream Transformerenhanced adaptive graph convolutional network for extracting the features on each joint that can be used to identify thermal adaptation behavior to reduce the influence of noise or incomplete skeleton on recognition results and to improve the robustness of the model.In summary,the development of artificial intelligence-driven smart buildings is a major initiative to achieve a comfortable building environment and reduce energy consumption.And the scientific problems and key technologies involved in the optimal decision making of indoor thermal environment operation of buildings with human in the loop are extremely challenging.In this paper,we start from the practical needs faced by thermal adaptation behavior detection,and design models and algorithms based on the data characteristics and problems to be solved to achieve non-invasive detection of indoor multi-person thermal sensation.This topic belongs to the intersection of computer,architecture and other multidisciplinary frontier directions,and has significant theoretical and application values for the development of a new generation of artificial intelligence-driven intelligent building systems with independent intellectual property rights.