基于梯度下降和角点检测的玉米根茎定位导航线提取方法

针对玉米根茎图像信息，提出一种在拔节期后玉米大田环境下快速、精准提取导航基准线的新方法。首先利用2G-B-R和最大类间方差法分割图像，并利用形态学处理提高图像质量，对去噪图像像素按列累加获取垂直投影。传统峰值点法在寻找特征点时需要设定阈值，耗时长且伪特征点多，因此提出一种基于梯度下降的特征点寻找方法，利用某点沿梯度下降的方向求解极小值从而求得特征点。根据角点检测原理，利用特征点像素各个方向梯度变化不同剔除伪特征点，解决了传统算法异常点过多、错误剔除玉米根茎定位点等问题，最终采用随机采样一致算法拟合导航线。试验结果表明，与传统算法相比该算法能够很好的适应复杂环境，实时性强，即使在缺苗、杂草等情况下仍具有很强的鲁棒性，平均处理准确率为92.2%，处理一帧分辨率为1 280像素×720像素的图像平均耗时为215.7 ms，该算法为智能农业化机械在玉米田间行走提供了可靠的、实时的导航路径。

Extracting navigation line for rhizome location using gradient descent and corner detection

Here, a quick and accurate extraction was proposed for the navigation baselines of corn rhizomes using gradient descent and corner detection. The images of corn rhizomes were also captured at the jointing leaf stage in the field. First, the images were segmented by the 2G-B-R and the maximum between-class variance. The morphological processing was also implemented to improve the image quality. The vertical projection was then obtained to accumulate the denoised image pixels by column. Second, a feature point search method was also proposed using gradient descent. As such, the time-consuming and false feature points were avoided in the traditional peak point method. The initial point position was determined after smoothing the curve. Subsequently, the negative gradient direction of the position was used as the search direction. This direction was the fastest descending direction of the current position, where the minimum value position was found the fastest. The gradient of the mean square error loss function was also used as the descending step size. Iteratively, the location of the minimum value was achieved during this time. Therefore, the improved method enhanced the searching efficiency for the fewer errors of feature point search than the traditional. Third, the correct positioning point of the corn rhizome was retained to remove the most pseudo feature points using the corner detection. Specifically, each feature point was traversed to compare the gradient change of the pixel value in each direction of the feature point. The reason was that there were too many pseudo-feature points in the traditional searching, leading to incorrectly eliminating the positioning points of corn rhizomes. Finally, random sampling was used to fit the straight line, in order to reduce the interference of the false feature points from the rest. Since the center of the window was a corner point, the grayscale change of the point was the largest before and after moving. The larger weight coefficient demonstrated that the point contributed much more to the grayscale change, as the window moved. Once the grayscale changes of points farther from the center of the window were gradually approaching smooth, the weight coefficients of these points were set to be smaller, indicating that the point contributed much less to the grayscale changes. Consequently, several windows were moved in the surrounding direction, when the window function detected that there was no grayscale change around. Once the window was close to the edge feature line, the window was moved along the edge direction. The window reached the corner to realize the detection, as the surrounding grayscale changed greatly. Anyway, a random sampling was consistent to fit the navigation baseline. Different data sets were repeatedly selected to verify the model, thus iterating until a better model was estimated. The angle bisector of two navigation reference lines served as the cornfield navigation line. The test was performed on an industrial computer with the Intel Pentium(R) CPU G3250 @ 3.20GHZ, 4GB memory, and Windows 7 (64-bit) operating system under the integrated development environment of Python3.7 (Anaconda). Experimental results show that the improved model can be expected to better serve complex environments, indicating strong real-time performance, and the robust even in the absence of seedlings and weeds, compared with the traditional. The average processing accuracy rate was up to 92.2%. An average of 215.7 ms was used to process an image with a resolution of 1280 pixels×720 pixels. This finding can provide a reliable and real-time navigation path for the intelligent agricultural machinery in the corn field.