dc.description.abstract | In this study, we use structured light to measure three-dimensional information of objects. Traditional structured light project sinusoidal fringe structured light onto the objects, capturing images with CCD cameras, and measure their three-dimensional structure with a phase shift technology. In order to achieve the purpose of fast measurement, we use the three-phase shift technology and CUDA (Compute Unified Device Architecture) parallel processing algorithms to improve measurement speed. Further, for future application of continuous motion production line, we derive an arbitrary fixed phase difference of the three-step phase shift algorithm. Via projecting a fixed sinusoidal fringes structured light, we convert the moving speed of the production line itself to the relative phase shift amount so as to measure the height of the object.
Since three-step phase shift method takes less images, the phase error becomes more sensitive. Thus, we use computer simulation to generate the error data, then tabulate them for the follow-up experiments. Because this study uses a digital projector to project stripe, so we build relationships between the CCD camera capture and the non-linear projection of the projector. Through the relationship we calibrate the captured images to correct the phase error caused by nonlinear projection. For future application of motion production line, moreover, we have to remove the phase errors of objects caused by different degrees of distortion from different positions on the images. We first calibrate the camera with a checkerboard plate, and use intrinsic and extrinsic parameters as well as distortion coefficient to restore the distorted image to undistorted one. By this way we increase the accuracy of measurement.
After actual measurement, our accuracy can achieve millimeter, but we still have to conquer the large amount of error. We apply CUDA parallel computing algorithms to the three-phase shift technology and the average processing time is about 4.9ms. Compared with calculation in CPU, the performance is 38.7 times faster. If we consider the memory access time from memory to GPU, the overall performance increased roughly 3.3 times.
| en_US |