Research On Non-contact Human Thermal Comfort Detection Algorithm Based On Deep Learning

Recently,with the continuous development of the global economy,the level of energy consumption is also increasing.The traditional temperature control scheme of buildings causes a certain waste of energy and neglects the human body's feeling of cold and heat(such as feeling very cold or very hot when the air conditioner is turned on).The construction of real-time non-contact human thermal comfort detection can effectively alleviate the above situation,and realize the "peopleoriented" intelligent building.However,the current non-contact thermal comfort detection mainly uses infrared and other devices.Due to its high price and inconvenient installation,it cannot be well applied to the thermal comfort environment of intelligent buildings.Used at the same time,some of the existing image capture devices such as the related research is not very good accurate projections for the human body thermal comfort condition,therefore,the core section to solve above problems,this paper is based on deep learning and openpose,build up the related algorithms,the main research content is as follows:From the perspective of micro skin texture,skin texture images and the actual skin temperatureskin,image saturation(H)of the channel to the skin in the frequency domain are analyzed,two types of skin temperature prediction algorithm is proposed: the lateral prediction algorithm(individual)between the skin temperature and longitudinal(individual)skin temperature prediction algorithm:(1)the transverse(individual)between skin temperature prediction algorithm using the data of one part person for training algorithm model,predict the skin temperature of other individuals.Transverse algorithm using four different network structure,network ? only use RGB images,the three kinds of network structure is used by the H channel frequency domain information as an additional auxiliary input,the three kinds of network structure of H channel spectrum respectively the way to join the network made different attempts.The experimental results show that the error of the horizontal temperature prediction algorithm is a little larger when predicting the temperature of others.(2)Longitudinal(within the individual)skin temperature prediction algorithm uses the skin texture and truth data of the individual in the first half of the time to train the algorithm model and predict the skin temperature in the second half of the individual.The vertical algorithm uses three different network structures.network ? use RGB image,only after both the network use the H of data channel itself as an additional input,and try different ways of access.The experimental results show that the longitudinal temperature prediction algorithm can maintain a relatively ideal error when predicting individual skin temperature.Two kinds of microscopic skin temperature prediction algorithms can give the human skin temperature within a certain error range based on the different needs between and within individuals.From the perspective of macroscopic human body's hot and cold posture,the Openpose key data of 6 kinds of hot and cold posture were collected.(1)A static attitude recognition algorithm is proposed for the form of input data,but the static algorithm has some defects.(2)Therefore,the collected data were improved based on the action sequence,and the key frame data with characterization function in different stages of each action process were selected respectively,and combined into new sample data.(3)based on sequence data sets is proposed three kinds of gesture recognition algorithm,using the full connection,convolution structure respectively,network ? also according to different individual differences in the location to location data normalization processing,adds additional input layer and the network structure to receive the normalized data.The final results show that the action sequence data set has unique advantages.Convolutional structure is more effective in processing coordinate data of key points.Normalization can improve the robustness of the algorithm better.The experimental results show that the macroscopic cold and hot posture judgment algorithm can accurately judge the cold and hot state of human body when the body is in the thermal uncomfortable state of too cold or too hot due to the body's subliminal adjustment to make some actions.