基于DinoPose的列车司机手比行为检测研究

Research on Hand Signal Behavior Detection of Train Driver Based on DinoPose

摘要: 本研究针对铁路场景列车驾驶室驾驶员监控视频图像提出了一种列车司机手势动作识别算法模型DinoPose。通过引入Transformers中的编码器–解码器结构来实现基于回归的人体骨架关键点检测，有效地将Dino网络的应用场景从目标检测扩展至人体骨架检测。通过多组列车驾驶室的视频图像所抽取的关键帧数据集测试，本文提出法在精度上优于Openpose和Yolo-pose算法，其中mAP达到了95.72%，手比项点的检测准确率达到85.74%以上，能够满足铁路局机务段机车司机室监控视频智能分析的实际业务需求。

Abstract: This study proposes a train driver gesture action recognition algorithm model DinoPose for video images of train cab driver monitoring in railroad scenes. By introducing the encoder-decoder structure in Transformers to achieve regression-based human skeleton key point detection, the application scenario of Dino network is effectively extended from target detection to human skeleton detection. Tested by the key frame dataset extracted from multiple sets of video images of train cabs, the proposed method in this paper outperforms Openpose and Yolo-pose algorithms in terms of accuracy, where the mAP reaches 95.72% and the detection accuracy of hand ratio item points reaches more than 85.74%, which can meet the actual business requirements of intelligent analysis of locomotive driver’s cab monitoring video in the locomotive section of railroad bureau.