Skeleton Keypoints And Deep Reinforcement Learning Based Human Behavior Recognition Models

With the rapid popularization of video acquisition equipment and fast development of internet technology in recent years,video information has been increasing exponentially increase,which brings great challenges on manual analysis of these large amount of data.Even if a large amount of manpower and material resources are invested in mass video information analysis,missed detections and wrong detections are still regularly seen due to human’s own weaknesses.Therefore,it is necessary to develop an intelligent model to automatically analyze and recognize human behaviors.Considering the shortcomings of the deep learning,this paper proposes a skeleton keypoints and LSTM based two-layer bidirectional Seq2 Seq behavior recognition method(SB2＿Seq2Seq)and a skeleton keypoints and deep reinforcement learning algorithm based human behavior recognition method(SDRL).With the development of computer vision and deep learning,deep learning methods for human behavior recognition have been applied in many fields,such as smart monitoring,smart home,and medical diagnosis.Deep learning has brought great progress in human behavior recognition due to its strong perception ability.Convolutional neural network(CNN)has achieved good performance in image recognition,but they lack decision-making capabilities.In addition,the analysis and recognition rate of convolutional neural networks in time series data is low.Although the Long-Short Term Memory network(LSTM)can solve the analysis of time series data,it can not well understand the context information in sequence data.The specific research content of this paper includes the following two points:1.With reference to the the problem of CNN on time series data analysis and low recognition rate and the problem that the Long-Short Term Memory network can not well understand the context information of sequence data,a two-layer bidirectional Seq2 Seq behavior recognition model(SB2＿Seq2Seq)based on skeleton keypoints and LSTM is proposed.Firstly,the human behavioral skeleton images extracted from the videos are input into SB2＿Seq2Seq model.Secondly,three Seq2 Seq models with different structures are built for comparison with the proposed model in the experiments,which are also compared with traditional deep learning algorithms.2.With reference to the deep learning’s shortcoming of lack of decision-making ability,a human behavior recognition model based on skeleton and deep reinforcement learning algorithms is proposed.The method divides the human behavior videos into frames,where the skeleton images of the video frames are extracted.Then the obtained skeleton images are input into the deep reinforcement learning model for training.The environment will give the model a feedback signal as a reward or punishment.The model interacts with the environment to increase the score by ‘try and error’ process.When the final score reaches convergence,it finds the optimal model that can accurately recognize human behavior.Finally,the model is compared and analyzed with other traditional human behavior recognition algorithms on UCF50 dataset and UCF101 dataset.