基于机器视觉的散料装车控制系统研究

发布时间：2018-03-08 17:13

本文选题：散料装车　切入点：机器视觉　出处：《长安大学》2016年博士论文　论文类型：学位论文

【摘要】：散料装车广泛存在于多个工程领域,其中料位检测与车速控制是实现自动装车的关键环节。然而,现有装车系统中超声波测距料位检测方法存在着无法直接检测卸料口下方料位、不能快速得到完整料堆轮廓、检测结果易受外界环境干扰等不足,而连续装车控制则需根据车厢尺寸预先设定车辆移动速度。针对上述影响散料装车控制质量及系统可靠性的问题,本文首次提出基于机器视觉的散料快速装车控制系统,在作业图像中提取料门开度和料堆轮廓等关键信息用于控制装车料位与车辆移动速度,实现了连续装车方式下的散料自动装车。研究内容主要包括以下几方面:1.散料装车机器视觉控制系统设计及试验平台搭建。提出装车系统主次功能及其实现途径,说明仅以单个工业相机作为传感器时,通过对料堆轮廓曲线与料门开度的视觉检测即可实现连续装车料位控制。给出双闭环视觉反馈的散料装车控制系统原理与结构、视觉模块布置方式以及相应试验平台。结合系统结构及作业过程特点,确定了视觉系统的标定平面及二元二次多项式平面标定方法。最后,针对作业图像特点设计了料门开度模板匹配检测与料堆轮廓纹理识别检测算法方案。2.旋转及光照不变模板匹配料门开度实时检测算法。将GMM聚类模型参数作为圆投影变换中各圆环的匹配特征,解决均值圆投影匹配定位性能较差的问题。对GMM参数无法利用NCC实现光照不变性的问题,使用局部圆环像素值得到线性对比度拉伸公式系数,并用于更新标准模板GMM参数,与逐像素的处理过程相比计算量大幅减小。算法实时性方面,采用自适应降采样搜索策略来减少不必要的特征提取及匹配,结合查找表的EM迭代聚类仅遍历一次待匹配区域即可完成特征提取及匹配。最后,使用Kalman滤波对定位结果进行校正以提高可靠性。试验结果表明:所提算法与均值圆投影匹配算法相比特征相似度峰值更为显著、定位失效程度更小,对圆环数选取不敏感,各圆环数下平均定位误差小于5像素;在1010?搜索区域下边长30~100像素的矩形目标匹配速度小于30ms,约为均值圆投影匹配时间的2倍,可满足料位控制要求。3.光照鲁棒装车料堆轮廓实时检测算法。针对光照-反射模型类图像增强方法计算过程复杂的问题,采用不重叠分块局部二值化作为光照不均料堆图像预处理方法,料堆部分可取得均匀无块效应的二值纹理图像,且与其他背景物体纹理具有一定差异,二值化后仅有的2个灰度级也使共生矩阵特征提取计算量大幅减小。针对SVM子块识别结果中难以避免的误识别及料堆内部出现的大面积干扰,结合SVM后验概率及料堆位置约束对子块进行MRF图像分割,避免纹理特征建模并利用图像自身特点提高弱光及干扰情况下料堆轮廓检测的可靠程度。最后,在各列中对类别交界附近一定间距下的重叠子块进行纹理识别,选取后验概率变化最大的位置作为该列最终分割结果并拟合料堆轮廓曲线。试验结果表明:子块纹理特征SVM识别率为95.2%,MRF分割后的分割正确率为98.8%,料堆轮廓检测绝对误差平均为2.2像素、轮廓检测误差标准差平均为6.1像素,270600?像素大小的处理区域下本文算法每帧检测速度平均9.7ms,标准MRF-ICM算法每帧检测速度平均为20.7ms。4.车速可变的动态连续散料装车料位控制方法。针对现有散料装车控制系统需根据车厢尺寸预先设定车辆移动速度的问题,借鉴人工装车控制过程提出车辆移动速度可自动调节的改进连续装车控制方法。与已有散料连续装车控制方式相比,增加了车辆移动速度模糊控制环节,根据料位变化来协调控制并持续提高车辆速度及料门开度。设计了料门开度增量、车辆移动加速度及料门开度前馈增量的模糊控制规则及论域范围,内环料门开度控制采用PID控制器实现。试验表明:所提控制方法适用于车厢尺寸经常变化且不易被快速检测的场合,可适应不同高度的目标料位,车辆能够自动且较快的达到最大卸料能力对应的车速,而不需要对车速进行预先设定。试验平台下料位轮廓与目标料位相比平均误差为-6.2像素,误差标准差平均为4.3像素。
[Abstract]:Bulk loading are widespread in many engineering fields, including material position detection and speed control is the key to realize the automatic loading. However, there is not direct detection of the discharge port below the level of ultrasonic ranging material existing detection methods of loading system, can quickly complete data stack profile, results are vulnerable to the external environment interference, continuous loading control is required according to the preset size carriage vehicle speed. The effect of bulk material loading control system reliability and quality problems, this paper first proposed the fast loading of bulk material control system based on machine vision, image extraction operation in the opening of material door and heap outline key information for control material loading and vehicle speed, realizes the automatic loading of bulk material loading continuous mode. The research content mainly includes the following aspects: 1. the bulk material loading machine Design and test platform of visual control system design. Put forward primary and secondary functions loading system and its realization, that only with an industrial camera as the sensor, the pile profile with hopper door can realize continuous visual inspection of material loading control. Given the double closed loop of bulk material loading control system principle and structure of visual feedback the visual module layout and the corresponding test platform. With the system structure and operation characteristics, to determine the calibration plane and the two plane of two quadratic polynomial calibration method of vision system. Finally, according to the operation diagram like the design characteristics of opening of material door template matching detection and windrow contour texture detection algorithm for rotation and.2. illumination invariant template matching opening of material door real-time detection algorithm. The parameters of the GMM clustering model as the matching feature of each ring circle projection transform, solving mean round investment Effect of matching the problem of poor performance. The parameters of the GMM cannot use NCC to realize the illumination invariance problem, the use of local rings to pixel linear contrast stretch coefficient formula, and used to update the GMM standard template parameters, compared with the process of pixel computation is greatly reduced. The real-time performance of the algorithm, using the adaptive down sampling search feature extraction and matching strategies to reduce unnecessary, combined with the EM iterative clustering look-up table is only one pass through the matching area to complete the feature extraction and matching. Finally, the use of Kalman filter on the positioning results in correction to improve the reliability. The test results show that the proposed algorithm and mean circular projection matching algorithm feature similarity peak more significantly, the positioning failure degree is smaller, less sensitive to the number of ring selection, each ring number average positioning error is less than 5 pixels; in the search area below 1010? Long rectangular target 30~100 pixel matching speed is less than 30ms, is about 2 times the average circular projection matching time, which can meet the requirements of the material level control.3. illumination robust Loading Pile contour real-time detection algorithms. Based on illumination reflectance model image enhancement method to calculate process complex problems, using block local binarization as the uneven illumination of windrow image preprocessing methods do not overlap, a value of two texture image can be obtained without uniform part of pile block effect, and has certain difference with other background texture, binarization after only 2 gray level co-occurrence matrix feature extraction also makes the calculation for error identification and greatly reduced. The material is difficult to avoid the SVM block recognition results in the large area inside the pile of interference, and the probability of material inspection after combined with SVM pile position constraints on block MRF image segmentation, texture feature to avoid the use of modeling and improve image characteristic Weak light and interference case material reliability heap contour detection. Finally, in each column of the categories under a certain distance near the junction of overlapping sub blocks for texture recognition, selection of posterior position probability biggest change as the final segmentation results and fitting the column pile contour curve. The experimental results show that the block texture features the recognition rate of SVM is 95.2%, the correct rate of MRF after the split split into 98.8%, stacking and contour detection of absolute error is 2.2 pixel average, standard deviation of the average contour detection error is 6.1 pixels, 270600? This algorithm processing area size in pixels per frame average detection speed of 9.7mS, the standard MRF-ICM algorithm for dynamic average detection speed per frame 20.7ms.4. variable speed continuous bulk material loading control method. According to the existing bulk material loading control system according to the preset size carriage speed of vehicles, vehicle control reference manual This process improved the speed of vehicles can be automatically adjusted for loading control method. With the existing bulk continuous loading control mode, increases the movement speed of the vehicle fuzzy controller, according to the material changes to coordinate control and continuously improve vehicle speed and opening of material door. The design of the opening of material door incremental fuzzy control rules and the range of movement of the vehicle acceleration and the opening of material door forward increment, the inner opening of material door control by PID controller. The experimental results show that the proposed control method is applicable to the carriage size changes frequently and not susceptible to rapid detection occasions, can adapt to the target material of different heights, the vehicle can automatically and rapidly the maximum discharge capacity of the corresponding speed, without the need for speed preset. Test platform level contour and target level compared to the average error is -6.2 pixels, the standard deviation of the average error It's 4.3 pixels.

【学位授予单位】：长安大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TH24;TP391.41

【相似文献】