The Research Of Embedded Measurement System Based On Binocular Vision

Binocular vision measurement model is a commonly used visual model for the tridimensional measurement, reconstruction of three-dimensional, product testing, traffic navigation and others. In order to improve the real-time property and reduce the cost, in this paper a binocular vision system is built on the resource constrained ARM embedded development board, and a series of studies on the algorithms and computing efficiency of camera calibration, image pre-processing, image correction, stereo matching, object recognition, and depth calculations are done which carries on the attempt to the practical application and portable of binocular vision and achieved good results.The paper starts from the research of binocular stereo vision model, and introduces the binocular stereo vision linear model and distortion model, then select the forward parallel model of binocular vision and triangulation method for measurement. When building the experiment platform, the paper chooses the general ARM embedded development board OK6410 as hardware platform, the embedded Linux 3.0.1 operating system as software development platform and ordinary CMOS camera OV9650 as image acquisition equipment. Then paper designs the corresponding external connection electric circuit and compiles the corresponding embedded Linux platform’s driver.In camera calibration stage, firstly camera calibration methods are briefly introduced, and then Zhang Zhengyou’s calibration method is emphatically desribed, finally Matlab and OpenCV are used to calibrate the cameras and the detail errors are analyzed. The calibrated re-projection error is within 0.4 pixels, which belongs to the sub-pixel level and accords with system’s accuracy requirement.In image preprocessing and correction stage, a variety of commonly used algorithms for image filtering are analyzed and compared in this paper, and Gauss filtering algorithm is selected to preprocess the image. The collected image distortion errors are compensated, so that the image can reflect the real reality scene better. Bouguet stereo rectification algorithm is used for three-dimensional correction, so that the image pairs can meet "the epipolar constraint" which can reduce the complexity of the subsequent stereo matching.In stereo matching stage, this paper puts forward a quasi-dense matching algorithm using objects’longitudinal edge points as the feature points. Some kinds of edge extraction algorithms are analyzed and compared in this article first, and the Canny operator is selected as extraction algorithm. The Canny operator is modified to extract only the longitudinal edge, and the process rate can be improved by about 35.56%. This modification can omit the transverse feature points and reduce the matching calculation amount, but impact on results little. SAD is used as similarity measure for stereo matching and the method scan the matching points twice based on the "uniqueness constraint" which can filter the the mismatch effectively. In order to reduce the large window’s calculation, SAD calculation window shape is modified with "米" shape instead of a rectangular window. A minimum SAD pruning is added while calculating the SAD to reduce redundant computation.In object recognition and depth calculation stage, the paper, based on the "continuity constraint", puts forward a quick object recognition method using adjacent parallax interval class to describe the objects. Object recognition time is only about 5 milliseconds which presents a high real time performance. The method obtains the parallax value according to the parallax interval weighted average, and uses triangulation method for depth calculation. The system’s accuracy error is under 2.17% in the 1 meter and the time for processing a pair of images is in 0.5 second.