基于ResAC-UNet网络的玉米作物行识别与导航线提取算法研究

Corn Crop Row Recognition and Navigation Line Extraction Algorithm Based on ResAC-UNet Network

摘要:

在复杂非结构化农田环境中，导航线精确提取对农机和农业机器人实现自主作业至关重要。农业环境中普遍存在光照多变、地形起伏和杂草干扰等挑战性因素，传统图像处理方法在适应性、准确性和实时性方面表现不佳，难以满足智慧农业的视觉导航需求。针对这些问题，本文提出一种基于改进UNet的ResAC-UNet深度学习网络模型。该模型采用ResNet-50网络替代传统UNet的编码器结构，增强特征提取能力。通过优化跳跃连接，提高了分割速度和实时响应能力。在网络的瓶颈部分引入ASPP模块，实现多尺度感受野建模，保持高分辨率特征的同时捕获更丰富的上下文信息。此外，模型整合CBAM，增强对作物行边界的精确感知，有效防止了关键特征信息的丢失，进一步提升了分割质量。在分割结果的基础上，采用行锚法和RANSAC算法实现了高精度导航线的提取与平滑处理。将获取的前视图像进行鸟瞰图转换，消除透视效应，生成保留ROI的作物行俯视图。试验结果表明，ResAC-UNet模型在精确率、平均像素精度、平均交并比和召回率方面分别达到99.23%、95.44%、85.23%和94.71%，性能优于当前主流的Segformer、DDRNet、HRNet及DeepLabV3+等分割网络，ResAC-UNet的平均推理时间为15.26ms，满足智能农机视觉导航识别的实时性需求。在相机的ROI区域内可提取3条导航线，中间导航线的最大角偏差仅为0.96°，最大像素偏差为4.3像素，实现了高质量导航线可靠提取。对比其他导航路径提取方法，本文方法具有更高的准确性和稳定性。

Abstract:

In complex unstructured farmland environments, accurate extraction of navigation lines is crucial for agricultural machinery and agricultural robots to achieve autonomous operation. Challenging factors such as variable lighting, undulating terrain, and weed interference that are common in agricultural environments, making traditional image processing methods perform poorly in terms of adaptability, accuracy, and real-time performance, it’s difficult to meet the visual navigation needs of smart agriculture. To address these issues, a ResAC-UNet deep learning network model was proposed based on an improved UNet. This model used the ResNet-50 network to replace the encoder structure of the traditional UNet to enhance feature extraction capabilities. The segmentation speed and real-time response capabilities were improved through optimized jump connections. The ASPP module was introduced in the bottleneck part of the network to achieve multi-scale receptive field modeling, while maintaining high-resolution features and capturing richer contextual information. In addition, the model integrate CBAM to enhance the accurate perception of crop row boundaries, effectively prevented the loss of key feature information, and further improved the segmentation quality. Based on the segmentation results, the row anchor method and RANSAC algorithm were used to achieve high-precision navigation line extraction and smoothing. The acquired front view image was converted into a bird’s-eye view to eliminate the perspective effect, and a top view of the crop row with ROI was generated and retained. The experimental results showed that the ResAC-UNet model achieved 99.23%, 95.44%, 85.23% and 94.71% in precision, MPA, MIoU and recall, respectively, which was better than the current mainstream segmentation networks such as Segformer, DDRNet, HRNet and DeepLabV3+. The average inference time of ResAC-UNet was 15.26ms, which met the real-time recognition requirements of intelligent agricultural machinery visual navigation. Three navigation lines can be extracted in the ROI area of the camera. The maximum angle error of the middle navigation line was only 0.96°, and the maximum pixel deviation was 4.3, which realized the reliable extraction of high-quality navigation lines. Compared with other navigation path extraction methods, the proposed method had higher accuracy and stability. The research result can provide an efficient and robust visual perception solution for the autonomous navigation of intelligent agricultural machinery in the field, which had certain practical value.