基于空间注意力和可变形卷积的无人机田间障碍物检测

UAV Field Obstacle Detection Based on Spatial Attention and Deformable Convolution

为了解决植保无人机作业时,传统田间障碍物识别方法依赖人工提取特征,计算耗时较长,难以实现在非结构化田间环境下实时作业识别的问题,提出一种优化的Mask R-CNN模型的非结构化农田障碍物实例分割方法。以ResNet-50残差网络为基础,将空间注意力(Spatial attention,SA)引入残差结构,聚焦跟踪目标的显著性表观特征并主动抑制噪声等无用特征的影响;引入可变形卷积(Deformable convolution,DCN),通过加入偏移量,增大感受野,提高模型的鲁棒性。构建包含农田典型障碍物的数据集,通过对比试验研究在ResNet残差网络结构中的不同阶段中加入空间注意力和可变形卷积时的模型性能差异。结果表明,与Mask R-CNN原型网络相比,在ResNet的阶段2、阶段3、阶段5加入空间注意力和可变形卷积后,改进Mask R-CNN的边界框(Bbox)和掩膜(Mask)的平均精度均值(mAP)分别从64.5%、56.9%提高到71.3%、62.3%。本文提出的改进Mask R-CNN可以很好地实现农田障碍物检测,可为植保无人机在非结构化农田环境下安全高效工作提供技术支撑。

In order to solve the problem that the traditional field obstacle recognition methods rely on manual feature extraction,long calculation time,and it's difficult to achieve real-time recognition in unstructured field environment,an optimized unstructured field obstacle instance segmentation method based on Mask R-CNN model was proposed.Firstly,an unstructured field obstacle dataset was constructed by aerial photography and network search.And then based on the ResNet-50 residual network,the spatial attention was introduced to focus on the significant apparent features of the tracking target,and the influence of useless features such as noise was suppressed.In addition,the deformable convolution was introduced into the structure of the ResNet-50 to add the offset,increase the receptive field and improve the robustness of the model.Comparative analysis was made by adding spatial attention and deformable convolution to different stages in the structure of ResNet-50.The results showed that compared with the original Mask R-CNN model,the mAP values of Bbox and Mask in Mask R-CNN improved by adding spatial attention and deformable convolution in Stage 2,Stage 3 and Stage 5 of the ResNet-50 were increased from 64.5% and 56.9% to 71.3% and 62.3%,respectively.The improved Mask R-CNN can well realize field obstacle detection and provide technical support for plant protection UAV to work safely and efficiently in unstructured field environment.