论文部分内容阅读
野外田间的主动摄像机视觉,尤其是对作物割过与未割过的高相似颜色表面进行实时识别与跟踪是一项极具挑战性的工作。提出了两种全新的快速分割方法,以用于农业机器人导航。其关键是基于多尺度特征提取,通过求取k-层行像素极值的加权均值来形成窄带兴趣区,以及基于相邻行像素两类特征证据增强与多证据模糊判别进行分割增强。提出了新的方法,分割出导向线能够自适应环境的一些变化。同时,本研究还提出了一种快速透视变换算方法和摄像机主姿态的一次性校正方法,能够在1ms内完成对分割导向线参数的透视投影变换,在0.5s内通过自校正获取相机的主姿态角。开发了一套对园艺草割过与未割过的边缘进行在线跟踪的分析软件。试验和相应的误差分析结果令人满意(160×120显示分辨率下,能在55ms内自主做出经透视投影变换的作业机理想移动方向决策,普通难度的相似颜色序列图像的分割误差被控制在了平均5%以内)。对最佳适应步法(BFS)做了改进,提出了多行最佳适应步法(MR-BFS),在不降低正确性的前提下使其分割速度提高了100%以上。通过折衷组合多行最佳适应步法(MR-BFS)与多证据模糊增强法(MEFE)进行在线试验,获得了160×120分辨率下8~9帧/秒的边缘自动跟踪性能。边缘跟踪试验显示:自然图像中的不同色块和阴影对其分割影响不大,能够快速输出导向跟踪参数。如果待分割纹理表面的颜色距离相对较远,还可采用本文新提出的颜色分量运算+颜色位屏蔽方法。该方法能在320×240分辨率下,在20~30ms内实现全帧的鲁棒分割,获得田间实时的多边缘跟踪性能。该方法避免了耗时的计算和人的操作介入,可被进一步应用于农业机器人的实际导航控制中。
Active field camera vision in the field, especially in crop cut and non-cut high-similarity color surface real-time identification and tracking is a very challenging task. Two new rapid segmentation methods are proposed for agricultural robot navigation. The key point is based on multi-scale feature extraction. The narrowband region of interest is obtained by calculating the weighted average of pixel extrema in k-layer rows, and the segmentation enhancement is based on two types of feature evidence enhancement and multi-evidence fuzzy judgment. A new method is proposed to separate out some changes that the guideline can adapt to the environment. At the same time, this study also proposed a one-time correction method of fast perspective transformation algorithm and camera main attitude, which can complete the perspective projection transformation of the parameters of the split guideline in 1ms and acquire the main camera by self-correction in 0.5s Gesture angle. Developed a set of analysis software for online tracking of cut and uncut edges of gardening grass. The experiment and the corresponding error analysis results are satisfactory (160 × 120 display resolution can be independently made within 55ms through the perspective projection transformation of the machine to move the direction of decision-making, the average difficulty of similar color sequence image segmentation error is controlled At an average of 5% or less). In this paper, the best adaptive walking method (BFS) is improved, and a multi-row best adaptive walking method (MR-BFS) is proposed to improve its segmentation speed by more than 100% without reducing the correctness. The edge auto-tracking performance of 8 ~ 9 frames / s at 160 × 120 resolution was obtained through online experiments using the compromise combination of MR-BFS and MEFE. Edge tracking experiments show that different color patches and shadows in natural images have little effect on their segmentation and can quickly output the tracking parameters. If the color of the surface of the to-be-separated texture is relatively far apart, the color component operation + color bit masking method proposed in this paper can also be adopted. This method can achieve full frame robust segmentation in 20 ~ 30ms with 320 × 240 resolution and get real-time multi-edge tracking performance in the field. The method avoids time-consuming calculation and human intervention, and can be further applied to the actual navigation control of agricultural robots.