基于多源视觉检测与自适应校正的玉米茎秆直径测量方法

Measurement Method of Maize Stalk Diameter Based on Multi-source Visual Detection and Adaptive Correction

摘要:

玉米茎秆直径是评估抗倒伏能力和整体植株健康状况的重要指标,然而,在田间环境下,叶片遮挡和相机拍摄角度偏差等复杂因素给玉米茎秆的准确检测和直径测量带来了巨大挑战。本研究提出了一种基于颜色和深度信息（RGB-D）的玉米茎秆直径估测方法,基于多源视觉检测数据建立改进玉米茎秆检测模型APCM-DETR,提高遮挡情况下视觉检测的可靠性。添加深度引导自适应边界框处理模块(Depth-guided adaptive bounding box,DABB),确保检测框内包含足够且完整的有效深度信息;提出基于点云的自适应校正直径估算模块(Point cloud-based adaptive diameter estimation,PADA),补偿由拍摄角度偏差引起的测量误差并实现茎秆直径的精确计算。通过引入基于APBottleneck构建的APCM骨干网络并融合卷积注意力模块CBAM,增强通道-空间联合特征建模能力,从而提升复杂田间环境下检测性能;DABB通过分析检测框边界处有效深度点比例,自适应调整边界位置,以提高检测框区域内深度信息的有效性;PADA通过拟合茎秆边界点云计算茎秆倾斜角,对检测框内点云进行角度补偿,并重建校正后的茎秆横截面凸包,实现玉米茎秆直径测量。试验结果表明,APCM-DETR检测模型精确率为88.2%,召回率为83.4%,F1分数为85.0%,模型内存占用量为54.6 MB,推理速度为121 f/s,平均精度均值为85.8%,与基线模型相比,内存占用量降低26.1 MB。玉米茎秆直径估计试验结果表明,R2为0.88,RMSE为6.19 mm。研究结果不仅在实地场景中提高了茎秆识别能力和拍摄角度出现偏差下的测量准确性,还为精准数字农业提高无损化实时分析效率提供了实用的理论基础。

Abstract:

Corn stalk diameter is an important indicator for evaluating lodging resistance and overall plant health. However,in field environments,complex factors such as leaf occlusion and camera shooting angle deviations pose significant challenges for accurate stalk detection and diameter measurement. A method for estimating corn stalk diameter was proposed based on color and depth information (RGB-D),established on a multi-source vision detection scheme: an improved corn stalk detection model,APCM-DETR,to enhance the reliability of computer vision detection under occlusion conditions. Adding depth-guided adaptive bounding box (DABB) processing module to ensure that the detection box contained sufficient and complete effective depth information,and a point cloud-based adaptive diameter estimation (PADA) module to compensate for measurement errors caused by shooting angle deviations and achieve accurate calculation of stalk diameter. APCM-DETR improved detection performance in complex field environments by introducing an APCM backbone network based on APBottleneck and integrating the convolutional block attention module (CBAM) to enhance channel-spatial joint feature modeling capabilities;DABB adaptively adjusted the boundary position by analyzing the proportion of effective depth points at the edges of the detection box,thereby improving the validity of depth information within the detection box region;PADA calculated the stalk tilt angle by fitting the stalk boundary point cloud,applied angle compensation to the point cloud within the detection box,and reconstructed the corrected cross-sectional convex hull of the stalk to achieve diameter measurement. Experimental results showed that: the APCM-DETR detection model achieved a precision of 88.2%,recall of 80.3%,F1 score of 85.0%,model memory usage of 54.6 MB,inference speed of 121 f/s,and mean average precision of 85.8%,indicating a reduction of 26.1 MB in memory usage compared with that of the baseline model. The corn stalk diameter estimation experiment yielded an R2 of 0.88 and RMSE of 6.19 mm. The research results not only improved stalk recognition capability in field scenarios and measurement accuracy under shooting angle deviations but also provided a practical theoretical foundation for enhancing the efficiency of non-destructive real-time analysis.