基于Gait-AVG的设施火龙果园喷施作业人员重识别方法

ReID for Spraying Workers in Facility Pitaya Orchard Based on Gait-AVG

摘要:

作业人员是设施农业智能化监管的核心。本研究以农药喷施作业为例，针对复杂环境下设施温室作业人员监测困难的问题，以设施火龙果园内使用背负式喷雾器进行喷施作业的人员为研究对象，提出了基于Gait-AVG的设施火龙果园喷施作业人员重识别（Re-identification, ReID）方法。模型以ResNet作为主干，通过时序池化层和水平池化金字塔后获得多类特征；引入了均值池化特征融合方法，以增强复杂场景下喷施作业人员的ReID结果，利用多尺度信息减少计算成本；融合使用Triplet Loss和Cross Entropy Loss损失函数进行模型训练，以提升喷施作业人员ReID模型性能和泛化能力。为验证所提出方法的有效性，在自建设施环境喷施作业数据集的基础上，实现了兼具保持样本特征一致性及有效实现作业分类的喷施作业人员ReID任务。所提出的算法在CASIA-B数据集上的测试结果表明，算法在正常行走、背包行走和换衣行走任务上取得的平均Rank-1准确率分别为96.55%、92.19%和79.47%。在设施环境喷施作业数据集上进行验证得到喷施作业人员ReID准确率为91.49%，在不同遮挡、拍摄角度和不同光照情况下的平均准确率分别为78.06%、97.50%和96.00%。结果表明，该方法能够用于识别和跟踪设施环境中的喷施作业人员并有效进行人员ReID任务，研究结果可有效提升设施火龙果园生产效率并为火龙果园智能化监测提供技术参考。

Abstract:

The agricultural workers are the core of intelligent supervision in facility agriculture. Using pesticide spraying operations as an example, the challenge of supervising personnel in complex environments such as greenhouse facilities was addressed. The focus was on personnel who use backpack sprayers for spraying operations in facility pitaya orchards, and a re-identification (ReID) method for facility pitaya orchard spraying personnel based on Gait-AVG was proposed. Multi-scale image feature extraction based on ResNet was achieved by the model, resulting in diverse features after undergoing temporal pooling and horizontal pooling pyramid. To bolster the efficacy of ReID in discerning sprayers amidst complex orchard settings, a mean pooling feature fusion technique was introduced. This method not only mitigated computational overhead but also leveraged multi-scale information to yield superior performance outcomes. Leveraging two distinct loss functions, namely Triplet Loss and Cross Entropy Loss, the training model was synthesized to bolster the monitoring and generalization capabilities pertinent to spraying behavior recognition. In order to substantiate the efficacy of the proposed methodology, a comprehensive facility environment spraying behavior dataset was curated, ensuring consistency in sample features and effective classification. The experimental evaluation of the proposed methodology on the CASIA-B dataset demonstrated compelling performance metrics: average Rank-1 accuracies of 96.55%, 92.19%, and 79.47% were attained for normal walking (NM), walking with backpack (BG), and walking with coat (CL) tasks, respectively. Notably, the proposed sprayer ReID method was validated in a production orchard, achieving a recognition accuracy of 91.49%. Furthermore, robustness tests under occlusion, variation in shooting angle, and diverse lighting conditions yield recognition accuracies of 78.06%, 97.50%, and 96.00%, respectively. The results indicated that this method can be used to identify and track personnel involved in spraying operations within facility environments. This study could effectively enhance the production efficiency of facility pitaya orchards and provide technical references for intelligent supervision of pitaya orchards.