Research On Human Action Detection And Prediction Method Based On CNN-LSTM For Human-Robot Collaboration

Human action detection and prediction is one of the representative tasks of action recognition in the researches of computer vision.These tasks based on temporal information also have huge potential value in human-robot interaction scenarios.However,human action is a relatively complex three-dimensional space signal.How to efficiently and accurately recognize actions in a complex environment and predict the future behavior of humans based on actions is a severe task.In the past ten years,due to the emergence of deep learning methods,human action detection and action prediction tasks have also made great progress.This paper proposed a method to extract the temporal and spatial features of skeleton sequences and designs a temporal action detection architecture and a human action prediction architecture respectively.According to project requirements,this article is divided into the following parts:The construction of human action data set for human-robot collaboration.Recently,most of the existing public data sets for human action are behaviors oriented to daily life,and there are few behaviors in the context of the human-robot collaboration industry.Secondly,few existing public data sets are oriented to temporal action detection tasks.Therefore,this paper constructs a human-robot collaboration-oriented human action data set to complete the tasks of temporal action detection and human action prediction.This paper proposed architecture for temporal action detection based on skeleton sequences.The temporal action detection model based on RGB colored video has a huge amount of calculation and will cause inaccuracy in recognition due to the interference of background factors.Its model is not robust to scene changes.This paper proposes a temporal action detection architecture based on skeleton sequence,which encodes the three-dimensional skeleton coordinates of the human body into RGB color map,to realize the conversion of temporal action detection task into a one-dimensional target detection task.This paper proposed a human action prediction architecture based on skeleton sequence.Human action prediction has a huge application prospect in the field of human-robot collaboration,which can effectively improve the efficiency and safety of human-robot collaboration.This paper proposes a human action prediction architecture based on recurrent neural network,which realizes the determination of the future frame position of the three-dimensional bone coordinates of the human body.