复杂背景下基于机器视觉的葡萄叶片检测与跟踪方法研究

发布时间：2017-12-28 08:33

本文关键词：复杂背景下基于机器视觉的葡萄叶片检测与跟踪方法研究　出处：《甘肃农业大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来酿酒葡萄的种植向集约化发展,大规模种植给生长状态的监测带来了极大的挑战,而人工检测效率低、工作强度大。葡萄叶片的健康状态在很大程度上能反映葡萄的生长状态,其作为视频监控的首要目标。叶片目标的实时检测和跟踪也是后续研究的基础,如基于机器视觉的叶片的分割、病害的识别和缺素诊断技术需要在复杂自然背景图像中准确检测出叶片目标,此外,叶片在生长且不断改变其位置、状态,为了判断葡萄生长状态的变化,对于已检测到的叶片需进行不断跟踪。本文首先对退化的叶片图像进行了颜色复原,然后采用了多角度可变形部件模型的叶片检测算法,结合颜色直方图描述外观特征,建立一种具有判别能力的目标跟踪模型,并能通过模型的学习克服跟踪中对目标相似区域的偏移,实现了对葡萄叶片运动的准确跟踪。在叶片图像颜色复原方面,针对自然条件下获取的视频图像受到灰尘干扰发生颜色失真问题,在大气散射模型的基础上建立了一种图像颜色复原模型,并通过暗元原理对模型参数进行了估计,恢复后的颜色分量接近叶片的原始颜色。在叶片检测方面,本文提出了改进的混合多角度可变形部件模型的叶片检测算法。首先,采用了G/R图像中提取HOG特征,并对特征向量进行了PCA降维处理,有效消除了光照和背景变化的影响。其次对G/R颜色特征图像采用可变形部件模型训练出了正面、侧面和背面3角度的叶片检测器,在多模型匹配过程中采用降低阈值参数和非极大值抑制的方式对产生叶片检测候选集合进行后期处理,检测结果的总体性能有所提高。试验结果表明,在自然条件下的叶片平均检测率为88.31%,平均误检率为8.73%,叶片检测准确性相对较高。在叶片跟踪方面,针对葡萄叶片的运动特点,本文采用基于判别模型和颜色特征的跟踪方法。为克服在目标相似区域被跟踪的可能性,建立了目标干扰区域判别模型,并将该模型与目标判别模型相结合;在跟踪定位过程将位置函数的最大值作为相邻帧的目标位置,并对该位置的目标采用了自适应阈值的方法进行尺度估计,实现了叶片的准确跟踪。试验结果表明,叶片跟踪的准确性也相对较高,其重叠率高达0.83,平均中心误差为17.33像素。最后,基于Matlab开发平台创建出可视化的人机交互界面,实现了监控视频的自动目标检测与跟踪,并能通过切换IP地址对不同的摄像头实现了操作。
[Abstract]:In recent years, the development of wine grapes has been developing to intensive development. Large-scale planting has brought great challenges to the monitoring of growth status, and the efficiency of artificial detection is low and the intensity of work is great. The healthy state of grape leaves to a great extent can reflect the state of grape growth, which is the primary target of video surveillance. Based on the real-time detection and tracking of the target is also leaves for further research, such as the identification of disease leaf segmentation, machine vision and the deficiency diagnosis technology needed in complex natural background image to accurately detect the target based on the blade, in addition, leaves in the growth and changing its position, status, changes in order to determine the grape growth state the leaves, has been detected for continuous tracking. Firstly, the color restoration of leaf image degradation, leaf detection algorithm and then using the multi angle deformable part model, combined with the color histogram to describe its appearance characteristics, establish a discriminatory target tracking model, and through the model of learning to overcome the similar region offset on the target tracking, to achieve accurate to track the movement of grape leaves. In the image restoration of leaf color, for the video image acquisition under natural conditions by the dust interference problem of color distortion, based on atmospheric scattering model is established on a color image restoration model, and the model parameters were estimated by dark element principle, the original color components recovered close to the color of leaves. In the aspect of blade detection, an improved hybrid multi angle deformable component model is proposed in this paper. First, the HOG feature is extracted from the G/R image, and the feature vector is treated with PCA reduction, which effectively eliminates the influence of illumination and background changes. The features in G/R color images using deformable part model training out of the front, side and back blade detector 3 angle, by decreasing the threshold parameter and non maximum suppression of the blade was processed to detect candidate sets in multi model matching process can improve the overall performance test results. The test results show that the average detection rate of the blade under natural conditions is 88.31%, the average false detection rate is 8.73%, and the accuracy of the blade detection is relatively high. In the aspect of blade tracking, the tracking method based on discriminant model and color feature is used in view of the characteristics of the motion of grape leaves. In order to overcome the possibility of tracking in the target area is similar, setting up target interference region discrimination model, and the model and target discrimination model combination; during the positioning process will function as the maximum position of adjacent frames in the target location, and the location of the target by using the adaptive threshold method for size estimation, implementation the accurate tracking of leaves. The experimental results show that the accuracy of the blade tracking is relatively high, the overlap rate is as high as 0.83, and the average center error is 17.33 pixels. Finally, a visual human-machine interaction interface is created based on the Matlab development platform, which realizes the automatic target detection and tracking of surveillance video, and realizes the operation of different cameras by switching IP addresses.
【学位授予单位】：甘肃农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S663.1;TP391.41

【参考文献】