| Structured light vision measurement is one of the important methods for non-contact depth acquisition.In recent years,related research and commercial products have been in full swing.Structured light technology uses a projector to replace one of the cameras in binocular vision,achieving a transition from passive vision to active vision.The projection mode of the projector determines the coding form of structured light,i.e.time coding,spatial coding or direct coding.Based on the symbol M array,this paper studies the spatial coding method applicable to dynamic scenes to achieve depth acquisition.Spatial coding has the advantages of fast,high efficiency,easy control,etc.The specific implementation includes two modules of coding and decoding.The structured optical coding determines the resolution of depth acquisition.The designed primitive type is related to the difficulty of decoding,and the decoding algorithm which directly affects the correctness of the recognition,the accuracy and robustness of the system.The specific research contents of this paper include:Firstly,an encoding algorithm based on M-array is studied.Based on the sequence generated by DeBrujin,the array is generated by folding.The size of the array is set by the relevant parameters generated by the sequence,combined with the specifications of the experimental hardware system,further determining the size of the pixel occupied by the primitive.The designed primitive symbols improve decoding resolution without increasing decoding difficulty,through proposed symbol density spectrum(SDS)principle,analyze the rationality of coding primitive pattern design and finally determine projection pattern.Secondly,a complete decoding algorithm is proposed,including image preprocessing,feature point detection,primitive symbol recognition,matching point pair acquisition,and error matching correction.Among them,the image preprocessing combined with the template form is processed by bilateral filtering and nonlinear contrast enhancement;according to the design of the coding template,the feature points defined in this paper are divided into primary feature points and secondary feature points,a proposed adaptive cross template and corner detection algorithm are respectively used to realize feature point detection.For the identification of primitive symbols,transfer learning is used in this paper.The DenseNet network structure is used to realize the distribution of symbol categories through fine-tuning.The proposed topology is used for window construction,based on window uniqueness,obtaining matching point pairs.At the same time,the topological structure is also used in symbol separation and error correction mechanisms;complete error correction mechanism also includes coplanar constraints and polar line constraints.Finally,based on the structured light calibration theory,combined with the coding and decoding algorithms proposed in this paper,the structured light system is calibrated and the experimental parameters are given.Combined with the calibration parameters and disparity information,the final depth acquisition experimental results are given to verify the feasibility and effectiveness of proposed encoding and decoding algorithms. |
