基于ST-GCN的空中交通管制员不安全行为识别

Unsafe behavior recognition of air traffic controllers based on ST-GCN

为预防和监督空中交通管制(ATC)工作中的违章行为,利用智能视频分析技术,研究适用于管制员坐姿工作的不安全行为识别模型。首先,分析管制员不安全工作行为的隐蔽性特征,总结5种典型管制员不安全行为,包括伸懒腰、瞌睡、低头入睡、歪头入睡和半躺入睡,并构建管制员不安全工作状态视频数据集(CUWS);其次,提出一种能描述管制员坐姿的骨架关键点拓展算法,基于时空图卷积网络(ST-GCN)搭建适用于管制员坐姿与腿部遮蔽情况下的不安全行为识别模型ATC-ST-GCN,并给出管制员不安全行为识别的工作流程;最后,利用CUWS数据集进行ATC-ST-GCN模型的训练和测试,并利用管制室实际监控视频开展验证试验。结果表明:该模型能够在有限验证数据集上实现5种典型不安全行为识别,准确率达到93.65%。试验结果证明该模型具有一定的科学性与有效性。

In order to prevent and supervise the violation behavior in ATC,an unsafe behavior recognition model suitable for controllers'sitting posture was studied by intelligent video analysis technology.Firstly,the hidden characteristics of unsafe working behaviors of air traffic controllers were analyzed,and five typical unsafe behaviors of air traffic controllers were summarized,including stretching,dozing,falling asleep,crooked asleep and half lying asleep,and the video data set of controllers unsafe working status(CUWS)was constructed.Secondly,a skeleton key point expansion algorithm that can describe the controller's sitting posture was proposed.Based on ST-GCN,an ATC-ST-GCN model suitable for sitting posture and legs hidden from view was built,and the workflow of the controller's unsafe behavior recognition was given.Finally,the ATC-ST-GCN model was trained and tested using the CUWS dataset,and the verification test was carried out using the actual surveillance video of the control room.The results show that the model can identify five typical unsafe behaviors on the limited validation dataset,and the accuracy rate reaches 93.65%.The experimental results show that the model is scientific and effective.