Research Of Recognition And Registration Based On Depth-image

Surface inspection plays an important role in the field of industrial inspection.Modern intelligent manufacturing puts forward requirements for automatic and intelligent detection of surface indentation detection.Automated detection is performed by imitating human eyes.At present,the automatic detection system is basically detected by the area array CCD camera imaging.There are certain limitations,such as the light intensity,the geometry and shape of the measured object will have a great impact on the detection of the amount of depression on the surface of the object.To solve this problem,this paper proposes to use the laser displacement sensor to scan the depth information of the rail bush surface to generate depth images.Through the current depth image recognition classification,the classification of the correct depth image and the same template image registration,to achieve the sleeve surface pressure judging the eligibility of the trap.Firstly,this paper designs the overall scheme of recognition and registration based on depth image,and realizes the collection and processing of depth information: it includes the conversion processing of depth information and gray information;the median filter,binarization,corrosion and expansion,etc.Learning processing;least squares fitting and other error compensation processing.Then the different recognition algorithms were researched and compared.Finally,the neural network recognition algorithm was selected to experiment.In this paper,a two-layer convolutional neural network is added to the basic multi-layer perceptron network to improve the accuracy of recognition.With 20,000 samples,the recognition accuracy is as high as 99.17%.Finally,the different registration algorithms are analyzed and compared.Finally,the initial registration and accurate registration are combined to perform the registration based on depth images.Among them,the initial registration includes: removal of invalid points,voxel filtering,calculation of eigenvalues and other point cloud pre-processing work and RANSAC initial registration algorithm;precision registration using kd-tree search strategy and ICP algorithm combined.The improved algorithm improves the running speed of the algorithm by more than 30% on the premise of ensuring the matching accuracyThe method proposed in this paper converts the high-precision one-dimensional contour scan information into a depth image,uses the image recognition method to identify the target,and uses the template to compare the depth information of the target to determine whether it is qualified.This method not only can detect depth information with high accuracy,but also does not need to know the relative positional relationship between the scanning line and the measured object.The experimental results prove the effectiveness of the method.