Non-contact Estimation Of Human Height And Weight Based On Multi-stage Neural Network And A Single RGB-D Images

Non-contact vision measurement is a traditional research field of computer vision and computer graphics.Due to the influence of different postures,clothes and positions,it is always a challenging problem to accurately estimate human height and weight from images.Based on a single depth image and a single RGB-D image,this paper constructs an effective intermediate representation through multi-stage neural network and the developing network,to realize automatic and active measurement of body height and weight under any clothes,position,and postures.In the height estimation method,this paper uses a 4-stage neural network and a novel intermediate representation to estimate the human body height from a single depth image.First,the human torso image is segmented from a single depth image using a fully convolutional neural network.In order to enhance the edge information,we plus the input of high-frequency information.Then the human torso image is further segmented into four parts: head,upper body,thigh and calf,and predict their lengths respectively as our intermediate representation of height estimation.Finally,a suitable network architecture is designed to estimate the human body height from above intermediate representation.In the weight estimation method,this paper uses 3 neural networks to form two modules to estimate the body weight from a single RGB-D image.In Module ?,two neural networks are constructed to remove human identity characteristics,and retain the clue of human posture and clothing thickness information to from intermediate representation of weight estimation.In module ??,a neural network is designed to estimate human body weight from the intermediate representation.This paper proposes a novel training method of neural networks called the developing network.By increasing or decreasing convolution layers,the fitting state of the neural network is destroyed repeatedly,so that the neural network jumps out of the local minimum and approaches the global minimum.The technique is applied to the final stage of height estimation and weight estimation.This paper constructs a human soft feature data set with 2136 RGB-D images.In this data set,each volunteer wears clothes in spring(autumn),summer,and winter.And they can make various postures within the range of the depth camera.In this data set,our accuracy of height and weight estimation reach 99.1% and 97.2% respectively.The experiments show that the accuracy of method is better than other existing methods.