基于改进可变形部件模型与判别模型的葡萄叶片检测与跟踪

为解决酿酒葡萄生长状态自动监测问题,该文提出基于机器视觉和视频处理技术的自动监测系统,开发了多角度可变形部件模型的葡萄叶片检测算法和基于颜色特征的判别模型跟踪算法。在叶片检测方面,该算法对颜色特征图像采用可变形部件模型训练出多角度叶片检测器,通过多模型匹配后产生叶片检测候选集合,选择集合中得分最高的检测结果作为最后的定位信息;在跟踪方面,结合图像中目标的颜色直方图,建立具有区分背景和目标的组合判别模型,并将位置函数的最大值作为相邻帧的目标位置,从而实现对叶片目标的实时跟踪。试验结果表明,该文检测算法对自然条件下的葡萄叶片平均检测率为88.31%,误检率为8.73%;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,其证明了该算法的有效性,研究结果为葡萄生长状态的自动分析提供参考。

Grape leaves detection and tracking based on improved deformable part model and discriminative model

Abstract: Recently, some researchers have exploited computer vision based video analysis to monitor the growth status of crop under the natural condition. Since leaf is the largest organ of the vast majority of plants, it often serves as primary monitoring object. Most of algorithms of illness analysis detect the blobs on a leaf surface and then judge which kind of diseases. In a leaf image, the blobs may be caused by shadow, dust, highlight, etc., which is prone to confuse with the blobs caused by diseases. To accurately judge the illness of a leaf for online surveillance, it is important to consider the time factor, since the blobs caused by the aforementioned factors may disappear with time elapsing. There exist some reasons such as the various poses, mutual occlusion, appearance and the irregular movement, which make the conventional detection and tracking methods hard to locate the leaves accurately in the images. In this paper, a novel scheme to monitor the leaves of vine grape was proposed. To improve the accuracy of leaves detection, a traditional RGB image was replaced by a G/R image to train the deformable part model (DPM) since the former makes it easier to distinguish the grape leaves and background than the latter. The DPM detector for leaves was constructed based on HOG feature, which was a mixture over three components representing a different aspect of a leaf. Since high dimension of HOG feature hampered real-time detection, PCA method was exploited to reduce its dimension, which speeded up the process of training and detection effectively. By utilizing the trained model, the overall score was computed for each root location according to the best possible placement of the parts through the matching procedure. The scores were sorted and the detection with respect to the highest score was picked out. To robustly trace the sharp movement of a leaf, probability model based online object tracking algorithm with color features was put forward. In proposed algorithm, object-background model capable of differentiating a leaf from the background was constructed firstly. To reduce the risk of drifting towards regions which exhibit similar appearance of leaf (but not really leaf) at a next frame, then a distractor-aware representation was combined to the formal model to generate a discriminative object model. Based on this model, overlapping candidate hypotheses densely was sampled within the search region and computed both the vote and distance scores for each candidate. This allows us to efficiently obtain the new object location in the next frame. In the long term tracking process, detection repeated at 30 minutes intervals to check whether new leaves appeared in the vision field. For the sake of the robustness, the images were gathered at various conditions, such as sunny, cloudy, shadow, flowering stages and fruiting stages, to train the detection and tracking models. Experiments were conducted to evaluate the performance of leaf detection at five different setting. The experimental results showed that the average detection rate reached up to 90.13%, and average false detection rate fell down to 8.52%. For the tracking algorithm, results also were exciting: the overlap rate was as high as 0.85, and the average center error was 17.33. Compared with the classical KLT tracking algorithm, our algorithm demonstrated the better robustness in the condition of illumination change and sharp movement.