基于异步交互聚合网络的港船作业区域人员异常行为识别

港船作业区域人员的异常行为识别可为智能航运的管控与决策提供重要数据支撑，有利于推动智慧港口和智能船舶的发展。基于异步交互聚合网络开展了面向港船工作环境下的人员异常行为识别研究。基于YOLO模型对港船图像进行卷积操作，利用特征金字塔优化卷积结果得到图像序列中每一帧的人员位置，结合联合学习检测和嵌入范式输出港船图像序列中的人、物体特征信息以及时序信息；利用异步交互聚合网络中的交互聚合结构更新特征池的多维度特征信息，以识别港区与船舶工作环境下的人员异常行为。实验结果表明：提出的港船作业区域人员异常行为识别方法的平均识别精度为91%，在港区工作环境下的人员异常行为识别精度为85%，在船舶驾驶台环境下，提出的异常行为识别框架对船员的不安全行为识别精度达到97%。所提出的识别框架在不同港船作业区域环境中都能获得较好的精度，验证了其有效性和可靠性。

Detecting Abnormal Behaviors of Workers at Ship Working Fields via Asynchronous Interaction Aggregation Network

Identify abnormal behaviors of workers at ship working fields provides important information for intelligent shipping management and decision-making, which is conducive to promoting the development of smart ports and intelligent ships. To achieve this, an abnormal behavior recognition framework is proposed based on a novel asynchronous interaction aggregation (AIA) model. The proposed model implements the convolution operation on the maritime surveillance videos by using the YOLO algorithm. The convolution results are optimized using the feature pyramid to locate the human in each image. A method of joint learning of detection and an embedding model are then integrated to extract the spatial-temporal features of workers and objects. Furthermore, the proposed AIA model utilizes an interaction aggregation module that update multi-dimensional feature information in the feature pool to detect abnormal behaviors of workers at ship working fields. The results show that the average recognition accuracy of the proposed method is 91%, and the recognition accuracy is 85% at the working fields. For the ship bridge monitoring, the recognition accuracy of unsafe behaviors of crews can reach up to 97%. Based on its validity and reliability, the proposed framework can achieve good accuracy in a variety of ship working fields.