Log Volume Detection Based On Binocular Stereo Vision

Binocular stereo vision is an important branch of computer vision, which could perceive the objective outside world by simulating the human visual system directly. Hence, it has widely applied in fields of 3D non-contact measurement, robot navigation, virtual reality and etc. Scaling of logs, especially in terms of log bundle, and log piles, are hardly to obtain exact results by traditional method such as weighing method, and visual plane measurement. Therefore, the technology of binocular stereo vision measurement of log piles has very important theoretical and practical significance. This dissertation discussed the key problems of binocular stereo vision systems, and proposed several camera calibration algorithms, matching algorithms and shadow elimination method. The main work of the dissertation is as follows:(1)Calibration method with two orthogonal vanishing points was proposed to solve the problem of large error and ill condition of common calibration based on vanishing points. This method requires only two images, firstly, two pairs of orthogonal vanishing point were extracted from each image; secondly, winner point and focal length on each single camera were solved by linear algorithm; thirdly, considering the radial lens distortion, distortion correction was implemented by linear constraints. During this process, the optimization differential evolution algorithm was proposed. Parameters within the solution were set as initial value. Line collinear features were set as constrained optimization parameters. According to the coordinates of 4 vanishing point, by using the infinite homography matrix rotation orthogonal constraint, the initial values of the rotation matrix can be obtained and more accurately calculated rotation matrix can be obtained by the iterative method. This calibration method only requires images with the camera template plane in more than two directions, which template can move freely without the motion parameters. Experimental results indicate that the reconstruction error is 0.598pixel. The experimental results show that the calibration method and can meet the precision requirements, and it is stable and reliable, high precision and strong robustness for self calibration.(2)Adaptive threshold merging method was proposed to solve the over segmentation and over merging problem caused by inappropriate meanshift segmentation parameter. In order to avoid over merging, the area size should be calculated first, and if the area size was greater than a certain threshold, then it would be judged as a single log end surface and do not take any merging. If not, calculate the distance EC between adjacent regions according to the pre-segmentation regions for the distance of couple areas in each image is different. Usually, the average distance of adjacent region was selected as threshold to realize adaptive merging. The experimental results showed that more accurate segmentation could realized in three different types which were the shadow area, the background and the end surface. The adaptive threshold improved the fixed parameter meanshift segmentation, and prevented over segmentation and over merging.(3)In the natural environment, the end surface of the log is not in a plane, and concave logs always fall into the shadows, which cause it is difficult to distinguish the gap between the logs. When logs were naturally stacked in the sunlight, in order to accurately segment the log ends surface and solve the problem of shadow elimination, the method of pattern recognition was proposed.30 dimensional color histogram and LBP texture histogram features were used in the classifier training and prediction. After image segmentation of end surface from stacked logs, this method is proved to be the effective in shadow image segmentation.(4)Stereo matching algorithm used a combination of FSAD (fast SAD) and meanshift segmentation was proposed to improve the stereo matching speed and accuracy of image segmentation. Firstly, extreme line calibrations were carried out in the left and right images using the camera parameters, and extraction of two pair of same points as constraints were done to complete the rectification, which can ensure the corrected image have the same ordinates at corresponding points. Secondly, FSAD algorithm was used to calculate the original disparity map. To extract a better image edge, meanshift image segmentation were taken on the original image, and got the final disparity by the optimization of the original disparity map. And the minimum enclosing matrix was used to calculate the end of long diameter and short diameter. The experimental results show that, the matching algorithm can obtain dense disparity map, and can obtain better matching accuracy at the edge of the regional disparity discontinuities, whose size measurement can verify the accuracy of the algorithm and can effectively solve the log end surface segmentation. The matching precision was improved by this method.Experimental results show that methods proposed in this dissertation meets the real time and precision requirements of log scaling, the processing time of a picture is less than 15 seconds, the average detection rate of root number is 96.17%, diameter measurement results is consist with manual measurement. Especially in the special environment such as unstable outdoor light, branches occlusion, and contaminated log ends, that methods provide a useful reference to the log visual measurement.