Research On Behavior Recognition Based On Network Of CNN And LSTM

As one of the most popular algorithms for machine learning,deep learning has solved many problems in the field of pattern recognition,especially in the field of image.It has achieved great results.With the rapid development of the era of big data,image and video data has exploded,and human behavior recognition has received extensive attention in numerous image studies.Human behavior recognition has great application value in various aspects such as intelligent monitoring,robots,intelligent medical systems,and smart homes.This paper is based on the convolutional neural network in deep learning to conduct research on behavior recognition.Since the depth image contains rich spatial information,this article is mainly based on the depth image for human behavior recognition research using convolutional neural networks and Bayesian optimization algorithms.Secondly,due to the lack of time feature description,this paper introduces long and short-term memory network combined with convolutional neural network to study the spatiotemporal features of behavior recognition.The specific content is as follows:Although many traditional behavior recognition methods have achieved good performance,these methods do not have clear model changes and interdependence in time.Convolutional neural network is a typical deep learning model that can achieve excellent performance in behavior recognition without relying on manual feature extraction.Based on deep learning,this paper studies a convolutional neural network that uses deep image sequences to recognize human behavior.First,we designed a simple three-dimensional convolutional neural network,which directly learns spatiotemporal features from the original depth image sequence.Using Bayesian algorithm to optimize hyperparameters,the experimental results show that the designed three-dimensional CNNs model is effective for simple human behavior recognition and even complex interaction recognition,and produces a competitive classification performance.Secondly,the influence of different time periods on the accuracy of behavior recognition is analyzed,and it is proved that increasing the time period will improve the accuracy of behavior recognition.Finally,in order to verify the generalization ability of the training model,we conducted a migration learning study by transferring the learned features to other human behavior data sets.The experimental results show that the designed three-dimensional convolutional neural network structure has a good Generalization performance,and achieved a large performance improvement.Although most networks for skeleton-based human behavior recognition are based on convolutional neural networks,convolutional neural networks can only extract features of spatial information,and lose time series information.In response to this problem,this paper introduces the long and short-term memory network to extract the time characteristics.This article first improves the VGG16 network model,replacing the fully connected layer with a global average pooling layer to prevent overfitting and reduce training parameters.Then,a comparative experiment of the CNN-LSTM model and the LSTM-CNN model was carried out on the MSRC-12 behavior data set.Due to the many parameters of LSTM and the large amount of calculation,a simpler GRU model with the same effect is used instead of LSTM for comparison experiments.The experimental results show that GRU-CNN has the best effect,and the experimental results are better than other algorithms on the same data set.