Research And Application Of Three-dimensional Object Measurement Based On CNN

With the continuous development of information technology,various industries on the need of three-dimensional target non-contact automated measurement are gradually becoming stronger.In recent years,people have made extensive research on the measurement of three-dimensional targets,and put forward some accurate measurement methods.But the measurement process is mainly focus on processing depth information to obtain object's linear scale,the way to abstract characteristics of object,which can be used to measure three-dimensional target,is still a challenging subject.Combined with the depth image with the characteristics of three-dimensional target depth information and deep learning with the characteristics of automatic feature extraction,the subject propose three-dimensional target measurement program based on the deep learning.On the basis of expounding the principle of the conversion of the point cloud from the depth image deeply,this paper designs and implements the filtering algorithm of the background cloud based on the background difference method.The subject chooses the outlier filtering algorithms by comparing the filtering effect of the three outlier filtering algorithms,and introduces the point cloud segmentation method used in this paper.The above algorithm is used to collect and process the three-dimensional target,and the final target cloud is mapped to the normalized depth image.The normalized depth image can represent the scale and depth information of the target surface,and the subsequent three-dimensional target measurement laid the foundation.The algorithm is implemented by coding,and the target point cloud collection program and the point cloud conversion depth image program are developed respectively.In this paper,the principle of convolution neural network from depth learning is studied,and the characteristics of Alex Net network and the working principle of the classification layer are analyzed.The Caffe deep learning framework for convolution neural network is studied.On the basis of this,the paper analyzes the characteristics of the continuous label values used in this subject and the working principle of the activation function for the regression task.The regression layer is designed for the convolution neural network and the experimental is conducted to verify the CNN regression network's the linear scales andnonlinear scales feature extraction ability.Finally,a three-dimensional target measurement scheme based on depth learning is developed by combining the improved Alex Net network and normalized depth image acquisition algorithm,and the scheme is applied to pig weight.The target point cloud collection program was used to collect the point cloud of the live pig back,and then the point cloud of the pig was converted into a normalized depth image using the point cloud conversion depth image program.The depth image can represent the three-dimensional structure of pig back,and has a strong correlation with pig's weight.During a period of time,pig's data(back point cloud and weight)were collected several times a day to accumulate rich specimens.After the data acquisition,the acquired point cloud data of the continuous weight section is converted to the depth images,and then the depth images of the pig labeled by the weight are trained by the improved Alex Net network to extract the abstract characteristics related to weight,achieving the ultimate goal that using computer vision to measure three-dimensional target such as to measure pig's weight.