基于关键姿态的快递场景人-物交互行为识别方法

Human-object Interaction Behavior Recognition Method in Express Scene Based on Key Gesture

摘要:

开箱验视是邮局快递场景中的一个重要环节,为了防止包裹内存在易燃易爆等危险品,快递打包前工作人员需按照行业要求实施危险品开箱验视。在人体行为识别框架中引入目标检测和关键姿态估计算法,提出了基于深度学习的快递场景人-物交互行为识别算法。首先,通过改进高斯混合(GMM, Gaussian Mixture Model)背景建模方法检测运动目标,提取行为关键帧,采用OpenPose算法进行姿态估计,识别初始行为类别；其次,针对常规行为识别方法丢失物品语义信息的问题,使用YOLOv5算法检测感兴趣物体类别和位置,提出基于拍卖算法(Auction)的多人-多物最优分配算法,构建人-物交互关系特征描述子；最后,将初始行为标签和人-物交互关系描述子进行决策融合得到最终识别结果。以实际快递场景数据对所提方法进行验证分析,实验结果表明,该方法可以对相似目标和多人干扰的复杂环境中的开箱验视行为进行准确识别。

Abstract:

Unpacking inspection is an important part in the express scene. In order to prevent the presence of flammable and explosive dangerous goods in the package, the staff must carry out unpacking inspection of dangerous goods according to industry requirements before packaging. In this paper, target detection and key pose estimation algorithms are introduced into the human action recognition framework, and a deep learning-based human-object interaction behavior recognition algorithm is proposed. First, by improving the Gaussian mixture (GMM) background modeling method to detect moving objects, extract behavior keyframes, use OpenPose algorithm for pose estimation, and identify initial behavior categories; The algorithm detects the type and location of the object of interest, proposes a multi-person-multi-object optimal allocation algorithm based on the auction algorithm, and constructs the feature descriptor of the human-object interaction relationship; finally, the initial behavior label and the human-object interaction relationship are described. The sub-processes perform decision fusion to obtain the final recognition result. Taking the actual express scene as an example to verify and analyze the proposed method, the experimental results show that the method can accurately identify the unpacking inspection behavior in the complex environment with similar targets and multi-person interference.