Research On 3D Measurement Based On Encoded Structured Light

Encoded structured light 3D measurement is one of the main technologies employed in 3D measurement and 3D reconstruction,which has many advantages such as nondestructive,low cost,high precision,and requiring simple devices.However,there are some difficulties for 3D measurement based on structured light,such as designing coded pattern,feature matching,correctly decoding and the robustness of the method.Due to the uncertainty of surface color,texture,luminance and illumination,dynamic object has low temporal correlation and spatial correlation.Existing coding and decoding algorithms of spatial encoded structured light are prone to be affected by the surface color,texture and luminance,which will reduce the measurement accuracy,and even can't acquire 3D information.In order to solve such problem,a new pattern is designed based on pseudorandom theory,which is encoded by binary geometric symbols,more anti-interference to surface color,texture,high reflectance,and owns large coding capacity and small window size.At the same time,according to the special structure of the encoded pattern,a new feature detection algorithm is put forward,which is implemented by multiple templates.The experiments show that the algorithm has strong performance of anti-noise and shows good performance while detecting feature point of surface with different curvature.In order to reduce the error rate of feature detection in aberrant image,plane constraint and topology constraint are proposed based on the basis of the epipolar constraint.The restriction mechanism can improve the accuracy of feature detection.Corresponding to our encoding scheme,a new decoding method is proposed based on convolutional network,which can implement the decoding stage with high accuracy rapidly.The decoding stage is implemented by the convolutional network based on deep learning theory.Firstly,to train convolutional network,symbol samples of different target objects are collected in different light environments.The target objects have different surface color and texture.Then,the trained convolutional network is applied in recognizing the encoded symbols in the decoding stage.Thus the aberrant image modulated by the surface of target object can be decoded and the 3D information can beobtained.The decoding experiment shows that the accuracy of the trained convolutional network could reach 98.07% for testing set.The decoding method shows good performance for dynamic objects with surface color,shadow,complex texture,strong reflectance,as well as encoded symbol samples with different sizes.To verify the performance of multi-template feature point detecting method,the proposed decoding method,optimize mechanism,objects with different surface property are applied in the 3D measurement experiments.By analyzing the result of the feature detection,point cloud,depth map and reconstruction map,we can conclude that in the process of 3D measurement,the proposed coded pattern,feature detection algorithm and the decoding method show better measuring precision and robustness.It can be employed in high precision 3D measurement for static and dynamic objects.