Intelligent Recognition Platform Of Operator Standard Behavior

To identify and feedback the operation behavior can timely prevent improper operation and avoid major accidents.It is difficult to manual recognition and apply sensor recognition on a large scale due to the high cost,and the traditional image processing scheme can’t meet the requirements because of the poor recognition ability.Based on the excellent deep learning algorithm in the field of computer vision,this paper aims to realize a set of intelligent recognition platform of operator behavior specification,which can effectively recognize,intelligently remind and save cost.The platform has complete functions and strong generalization,and the identification function is added to eliminate the interference of irrelevant personnel in the operation site,which is of great significance in improving the identification efficiency and effect of operation behavior specification.In the aspect of dataset building,this paper establishes a professional person body part detection dataset for function implementation,and selects sit,stand,walk and turn in HMDB dataset as the standard actions to be identified.In the aspect of platform function realization,and the AP50 of detection frame reaches 90.5%;Using the Mask R-CNN algorithm on the COCO keypoint dataset,the AP50 of the keypoint of the human body reaches 86.6%;The recognition results of Mask R-CNN are sent to the input of Long Short-Term Memory(LSTM)algorithm for training,and the Precision of behavior recognition is 85.6%;the scheme of face feature matching and triple loss function is adopted on CASIA-WebFace dataset,and the Precision of identity recognition is 90.7%in the test dataset collected by the project.In the aspect of algorithm improvement,this paper realizes the behavior information extraction algorithm with face detection function through the improved scheme of generating pseudo tags,which combine multi-stage feature extraction process to speed up the efficiency of the algorithm.This paper also adds a random clipping optimization strategy to the human body part detection and behavior information extraction algorithm,which can save more than 20%of the training time and provide a great help for the fast iteration of the algorithm.