Research On Abnormal Behavior Identification Based On Long Short-term Memory Neural Network

Behavior recognition has always been a research hotspot in the fields of pattern recognition and computer vision.It has a wide application market and huge application value in intelligent video surveillance,video data content retrieval,virtual reality and so on.The study of behavioral recognition involves hotspots and difficult problems in the fields of digital image processing,signal processing,pattern recognition,etc.The research on it has a high theoretical research value.Different from the traditional behavior identification method,this paper adopts deep learning method to study the behavior recognition.The traditional behavior identification method is often the feature extraction as the core of the study,while the use of deep learning algorithm to a certain extent realizes the algorithm's own learning data features,and can classify the effective features learned.This reduces human intervention to some extent.This paper extracts the keyframes from the behavioral image sequences and represents any behavior as a time-sequential behavioral key sentence.For the timing of key sentences,this paper uses the Long Short Term Memory Network(LSTM),which is good at processing time series data,to classify the key sentences of the behavior and achieves the abnormal behavior recognition for the parking scene.Considering deep learning algorithms usually requires a lot of data to train a better model.In the case of existing data,this paper uses the methods such as the Deep Convolutional Generative Adversarial Networks(DCGAN)to generate data and achieve the effect of increasing the amount of data and balancing the differences between different types of data.In this paper,the CASIA behavior database of the Chinese Academy of Sciences and the extended behavioral database are verified experimentally.The experimental results show that the abnormal behavior recognition method in this paper is effective.The content of this article is as follows,(1)The database has been expanded on the basis of existing parking lot behavior data,and more than 300 pieces of behavioral video data have been added.(2)A dynamic time warping(DTW)behavior recognition statement characterization method proposed.First,the motion cycle curve is calculated for the noise-reduced behavior sequence.Then,through the motion cycle curve and DTW method,the behavioral keyframes are extracted,With reference to the semantic understanding method,the behavior key frame is represented as a series of behavioral key sentences,and the temporal characteristics of the behavior sequence are well preserved.(3)For the current lack of parking scene data and imbalance of different types of data,this paper uses data enhancement,generative adversarial networks(GAN),DCGAN and other methods to generate key sentences of the behavior,which satisfy the diversity of data.Adequate and diverse data is a powerful guarantee for the strong generalization ability of this behavior recognition model.(4)A method of identifying abnormal behavior based on LSTM is proposed.The key sentence of behavior is a kind of data with time-series characteristics.Based on the LSTM has a better processing ability for sequential data,this article uses LSTM to learn and classify key sentences of behavior.This paper carries out experimental verification and comparison on the CASIA database and the extended database.Experiments show that the proposed method is better than Recurrent Neural Network(RNN),Convolutional Neural Network(CNN),and Support Vector Machine(SVM).Support Vector Machine(SVM)and other methods have better recognition performance.