Research On SIFT Algorithm Of The 3D Reconstruction Based On SOC

Computer vision applications are increasingly widespread,as a major challenge for computer vision,3D reconstruction makes two-dimensional images into 3D virtual scene simulation through the calculation and analysis.The speed of reconstruction is relatively fast.The shape of the object is not restricted.Now,3D reconstruction has been fully automated modeling,widely used in Industrial automation,graphics and image recognition,robot navigation system,medical imaging,defense aerial surveyingand some other fields.The rapid development of 3D reconstruction will completely change the life style of.Therefore,the research on the 3D reconstruction algorithm has great theoretical and practical significance.Now,most of the 3D reconstruction use computer software to achieve.but its computation is large and long time spent.The quality of the reconstruction needs to be improved.Above all,it is not applicable for scene reconstruction with few feature points.To tackle the problem,based on SOC,the system uses ZedBoard as the hardware platform to design and implement feature extraction algorithm of the 3D reconstruction.The development board integrates FPGA chip with dual core ARM.The software and hardware co-design not only retains the flexibility of software design,but also improves the overall system operating rate through the hardware parallel operation.The result is more ideal.The thesis analyzes and compares the 3D reconstruction algorithm,deeply research and design to achieve the SIFT algorithm.SIFT has the advantages of scale invariance,multiplicity,uniqueness.In this thesis,the system mainly includes Gaussian filtering module,DOG scale space forming module,key point detecting module,PL and PS transmission module and 3D reconstruction descriptor form module.According to the speed demand of the system,the hardware and software are separated.Because the scale space formation and the key point detection module take a long time,its are realized by hardware.without affecting the image feature point detection results,and taking into account the hardware resources,DOG scale space formation module is optimized.The system fully use the FPGA pipeline structure,and improve the efficiency of the LUT.In the transmission of PL and PS,using AXI-Stream mode and AXI-HP with high-speed transmission interface,the system design IP core to realize high data transmission.Finally,accrording to the a acquired key points of information,size of the key points and the gradient direction are extracted.3D reconstruction descriptor is formed on the PS.In addition,the thesis has completed the transplantation of Linux system and OpenCV on the Zynq.The corresponding test has been completed.The system has completed the extraction of image feature points.The 3D reconstruction descriptor is formed,which decreases the computation of the whole algorithm and greatly shortens the running time of the system.