A Study On Symbol M-Array Based 3D Detection Approach Utilizing Binary Structured Light

Machine has been making great contribution human productivity. But the efficiencyand inelegancy of most currently developed intelligent machines, especially autonomousmachines in open environment, are still expected to be improved greatly. One of mainshortages isthat we still lack of active and effective machine understanding approaches.All movement strategies might be wrong while unknowing working scene or targetstimely. And obviously, this machine understanding should be based on real-time detectingdata.Spatial geometry parameter is one of most important parameters for machineunderstanding, since spatial information of working scene can be utilized directly inobstacle avoidance, path planning and movement strategies, and 3D information of anobject surface can provide basic data for 3D description, reverse engineer and structurebased recognition of targets.In the first chapter of this paper, both domestic and aboard approaches forthree-dimension measurement, such as hardware based distance meters, machine visionbased depth detection methods and structured light technique, were firstly reviewed andsummarized. Advantages and drawbacks of each method were discussed in detail. It waspointed out that structured light technique is one of most perfect approaches forthree-dimension detection. And simultaneity, researched codification strategies andpatterns were discussed in detail. Aimed at the limitation of working scene of currentlydeveloped structured light technique, the idea of laser projection and optical filtering wasreferenced. Concerning the reduction of coding approaches and codewords while laserprojection was adopted, and the requirement of huge codewords for dynamic scenedetection, an innovation of symbol M-matrix based pattern was proposed in this paper.The proposed pattern was coded in both axes and all codewords were globe unique,suitable for the detection of dynamic scene. And the proposed pattern was also adoptablefor binary projection (including laser) since it utilized binary symbol as M-matrixelements.In chapter two of this paper, system model based on pin-hole was established.Transfer relationship among source coordination, image Coordination and globecoordination were investigated. Solutions for object depth retrieval and pixel movingroutes under different object depth were presented. Influences of system parameters, suchas the bias set of source and receiver, angles of source plane and image plane, distancebetween source and source plane, distance between receiver and image plane and objectdepth, on captured images were discussed based on pin-hole system model. Partial ofachievement of former research of mini-grid codification pattern were also presented.Score square broadening method in CCD image scanning, seeds algorithm, connectionoptimizing strategy and sub-pattern decodification method were discussed. Mini-gridcodification pattern is also a binary pattern, coded in both axes and all codifications are globe unique. Although mini-grid codification pattern was no longer utilized in this papersince sub-patterns were likely to break on discontinuous object surface due to the largesize of sub-pattern, but it still provide subsequent research with helpful ideas andrealization methods.Codification symbols were investigated and discussed in chapter three of this paper.Based on parallel thinning algorithm and simplified thinning algorithm, symbol skeletonretrieval was firstly implemented. Index pattern between neighbor code and crossingdefinition were established. Symbol traversal principle was investigated and traversalsequence was obtained. Since symbol recognizability and positioning accuracy ofcharacteristic points were important indexes for symbol selection, multiple-leveldecomposing method of traversal sequence and structurization method of each levelsequence were then investigated in this chapter. Results showed the proposed structuredsequence based recognition method was effective and robust to severely deformedsymbols even rotated symbols, And simultaneity, multi-span resolution based chain anglesequence was put forward. It provided high sensitivity and quality for the detection ofcharacteristic points in symbols, and also provided subsequent investigation withfundamental denotation. Ten symbols used as M-matrix elements were put forward at theend of this chapter.Generation method for large size M-matrix was investigated in chapter four of thispaper. Utilizing a n-order shift register as the feedback function, generation algorithm forq-elements n-order de Bruijn sequences was firstly discussed. And generation method forall equivalent sequences from one known q-elements n-order sequence was also provided.Broadening principles for generating two-dimension M-matrix from one-dimensionsequence via diagonal rule and such other principles were investigated, and investigationinstances showed that globe unique principle could not be ensured for all gap windows. Aso called "pieces growth" algorithm for generation of large size M-matrix was proposedand complemented in this paper. The proposed algorithm greatly reduced the range ofpossible combination of M-matrix and improved the generation performances byanalyzing interface allocations and growing principles of pieces. Growing sequence:piece ordering strategy and three piece-selection principle such as exports and importsbalanced principle, interfaces number balanced consumption principle and elementsbalanced usage principle were also discussed in this chapter. About 4 hundreds millionsof binary M-matrix pattern, which had 9 element number, and 85×63 in pattern size, 2×2in window size, were generated. So far as author knows, they are largest patterns withmost element number among all related researches. Further more, the generated patternswere codified in both axes and all codewords were globe unique in patterns, suitable forthe detection of dynamic scene.Decodification method of symbol M-matrix based codification images wasproposed in chapter five of this paper. Image pre-processing methods, such as imagefiltering methods and thresholding methods, were firstly investigated. Aiming at the proposed projecting pattern, pixel-percentage thresholding method, background thinningapproach and connection detection method for singular symbol retrieval, black noiseelimination method were investigated. Virtual end-branch recognition method and whitenoise elimination algorithm were also discussed in this chapter. Internal chain branchesbased symbol recognition approach was provided. Chain angle curve and positioningalgorithms for characteristic points of sharp points, recession points and cross points wereinvestigated. Codewords retrieved algorithm was implemented by analyzing the relativeposition of characteristic points. Positioning algorithm for first set of key measurementpoints were discussed simultaneously and six first set of key measurement points in eachsymbol were retrieved utilizing the proposed algorithm. At the end of this chapter,window combination method and window match method based on voting mechanismwere discussed. Window position in projection pattern was identified utilizing exclusivityprinciple of windows. The absolute positions of all symbols in a window and allcharacteristic points in the projection pattern were ulteriorly identified.Chapter six was implemental and experimental part of this paper. Firstly in thischapter, an image analysis and processing platform ChOpenCV, based on Ch and IntelOpenCV was proposed. Interfaces of Ch space to C library, Ch and C spaceimplementation and the build method of Ch dynamic link library were discussed.Integration mechanism of Ch and OpenCV were present. ChOpenCV included salientfeatures both from Ch and OpenCV, and provided subsequent research with an excellentplatform which was interactive, pure C code and open source. The advantages andshortages of traditional calibration approaches such as non-linear optimizing method anddirect linear transfer method were then investigated, and then a novel system calibrationapproach called "two-step method" was proposed. The proposed method took lightsource system, CCD system and all optical system as black boxes, and regarded beamlight parameters in outter space as exclusive concerned variables. Comparing withtraditional calibration approaches, the proposed method was much simple and effectivesince it avoided the requirements of optical parameters. Based on the image processingplatform ChOpenCV and system calibration, measurement and reconstructionexperiments were carried out. The analysis to measurement results showed that theproposed approach got a maximum percentage error of 1.45% at an experimental area of1310mm in width, 983mm in height and 1940mm in depth. And reconstructionexperiments showed that the three-dimension reconstruction results providedfundamental structures and surface curves of targets.Conclusions and innovation points of this research work were presented in chapterseven of this paper. Simultaneously, the shortages of current research work and prospectsof future research were also presented.