The Monitoring System Of Operation Platform Based On Union Pose Estimation

At present,the pose estimation of each part of the human body is implemented as an independent task based on different networks.In many scenarios,there is need to recognize different parts of the human body,such as locating the key points of body,hands and face simultaneously.So different networks are required to locate the keypoints of each part respectively,which occupy large space and have complex network structure.On the other hand,using multiple networks will lead to a large change in the computation,which is not conducive to algorithm deployment on the chip or mobile with limited computation.Therefore,this paper proposes a unified pose estimation algorithm for human body,hand and face based on joint training.In order to meet the needs of multi-target detection task of human bosy,a human parts detection dataset on body,hand and face is established.Using Faster-RCNN detection framework to train on this dataset to achieve the detection of human body,hand and face in the picture.At the same time,adding FPN,using RoI-Align instead of RoI-Pooling,and using multi-scale training improve the detection performance of target in different sacles,especially small ones.Based on the relevance between keypoints detection tasks of human body,hand and face,a unified pose estimation algorithm of human body,hand and face is proposed.Drawing on Mask-RCNN framework,this paper designs the keypoints detection network of face and hand to build a unified pose estimation network for human body,hand and face.Through the research of multi-task joint training strategy,the training,test and optimization of the model are completed.And finally,the unified model achieves APkp of 62.6%on COCO keypoint validation,PCK of 0.909 on CMU hand pose validation,and normalized error of 4.68 on 300W face keypoint dataset,which is superior to single-task models.Finally,a union pose estimation platform of body,hand and face is built to realize the behavior analysis of workers or operators in the monitoring video,including face recognition,behavior analysis of operation,behavior analysis on hand operation,etc.,which is used for workers' wrong operation alarm and irregular behavior record.