基于语义分割的中后期玉米行间路径导航线检测

Interrow Path Navigation Line Detection of Maize in Middle and Late Period Based on Semantic Segmentation

摘要:

中后期玉米行间路径存在光照不足、遮挡等因素的干扰，不利于农业机器人自主作业时导航线的检测。针对此问题，本文提出一种基于改进Fast-SCNN语义分割模型的中后期玉米行间路径导航线检测算法。首先，针对目前路径语义分割模型在中后期玉米环境下边缘分割不够准确的问题，提出一种Edge-FastSCNN模型，在模型分支中引入本文提出的边缘提取模块（Edge extraction module, EEM）以获取准确的路径边界信息，并在模型中引入空间金字塔池化（Atrous spatial pyramid pooling， ASPP）模块以融合图像边界信息和深层特征。然后，基于模型预测的行间路径掩码，通过像素扫描法检测路径掩码左右边界点，通过加权平均法求得路径掩码中点。最终利用最小二乘法拟合导航线，实现中后期玉米行间路径导航线的检测。为验证所提出方法的性能，基于中后期玉米正常光照无遮挡、光照不足、阴影、杂草遮挡、叶片遮挡等5种环境，进行了模型性能对比实验和导航线检测实验。实验结果表明，模型平均交并比为97.90%，平均像素准确率为98.84%，准确率为99.39%，推理速度为63.0 f/s;模型在上述5种环境下的平均交并比为96.93%~98.01%，平均像素准确率为98.33%~99.03%，准确率为98.53%~99.12%;预测导航线与真实导航线在上述5种环境下的航向角偏差平均值为1.15°~3.16°，平均像素横向距离为1.89~3.41像素;导航线检测算法的单帧图像平均处理时间为90.04 ms。因此，本文提出的导航线检测算法满足中后期玉米行间路径导航任务的准确性和实时性要求。

Abstract:

The interrow path of maize in the middle and late stages is interfered by factors such as insufficient light and occlusion, which is not favorable to the detection of navigation lines during autonomous operation of agricultural robots. To address this problem, an algorithm based on the improved Fast-SCNN semantic segmentation model for detecting the navigation lines in the interrow path of maize in the mid-late stage was proposed. Firstly, to address the problem that the current path semantic segmentation model was not accurate enough for edge segmentation in the mid-late maize environment, an Edge-FastSCNN model was proposed, and the edge extraction module (EEM) proposed was introduced in the model branch to obtain accurate path boundary information, and spatial pyramid pooling was introduced into the model. Atrous spatial pyramid pooling (ASPP) module was introduced in the model to fuse the image boundary information and deep features. Then based on the interline path mask predicted by the model, the left and right boundary points of the path mask were detected by pixel scanning method, and the midpoint of the path mask was obtained by weighted average method. Finally, the least squares method was used to fit the navigation lines to achieve the detection of the mid- and late-stage maize interline path navigation lines. In order to verify the performance of the proposed method, model performance comparison experiments and navigation line detection experiments were conducted based on five environments such as normal light without shade, insufficient light, shadows, weeds shade, and leaf shade of maize in the middle and late stages. The experimental results showed that the average intersection and merger ratio of the model was 97.90%, the average pixel accuracy was 98.84%, the accuracy rate was 99.39%, and the inference speed was 63.0 f/s;the average intersection and merger ratio of the model in the five environments mentioned above was ranged from 96.93% to 98.01%, and the average pixel accuracy was ranged from 98.33% to 99.03%, and the accuracy rate was from 98.53% to 99.12%;the average value of heading angle deviation between the predicted navigation line and the real navigation line in the above five environments was 1.15°~3.16°, and the average pixel lateral distance was 1.89~ 3.41 pixels;the average processing time for a single-frame image of the navigation line detection algorithm was 90.04 ms. Therefore, the navigation line detection algorithm proposed met the mid- and late-stage maize interline path navigation task’s accuracy and real-time requirements.