Anomaly Detection And Behavior Recognition Based On Spatiotemporal Feature Learning

Video-based behavior recognition and anomaly detection are widely studied in the fields of security,medical care,education,and energy.Current anomaly detection and behavior recognition Due to the variable scene of the monitoring data,imaging instability,occlusion and other factors,the model is difficult to train,the behavior and abnormality detection accuracy is not high enough and the speed is slow.Aiming at these problems,this paper proposes a spatio-temporal anomaly representation learning model and a three-dimensional convolutional neural network behavior recognition model.For behavior recognition,the 3D convolutional neural network is used to directly predict the behavior.The model training and test experiments are carried out on the UCF101 dataset.The experimental results show that the proposed method achieves competitive results compared with similar methods.For the anomaly detection,the spatio-temporal anomaly representation learning model is firstly constructed to extract the temporal and spatial features of the video.Then,the multi-dimensional Gaussian model is used to discriminate and extract the anomaly features,and then the feature reverse direction is used to calculate the exact position of the anomalous behavior in the video image.The spatio-temporal anomaly representation learning model proposed in this paper does not require specific specialized training.The proposed method has the advantages of strong versatility,fast calculation speed and high detection precision.We validated the algorithm on the representative surveillance scene datasets of Subway and UCSDSped2.The method of this paper achieves the detection rate of 18 FPS under the condition of common computing resources,meeting the real-time requirements.Compared with the similar methods,the proposed method achieves competitive results in both the frame precision measurement and the pixel precision measurement.This paper focuses on the video anomaly detection and behavior recognition algorithm based on three-dimensional convolutional neural network,and designs a multi-scale behavior recognition model based on three-dimensional convolutional neural network.Based on the behavior recognition model,a spatio-temporal feature learning model is proposed for extraction.The anomalous behavior of a video characterizes features.The main contributions of this paper are as follows:1)A deep learning model based on 3D convolution is proposed.In order for the model to learn more scale information,we designed a multi-scale network structure in the identification network.Through the multi-scale structure,the model can learn more space-time features at different scales.These features have stronger discriminating ability and will These features are used for behavior recognition,and the effectiveness of the proposed method is verified on open public datasets.2)Based on the behavior recognition model of this paper,a neural network model for character learning of abnormal behavior is proposed.The model extracts the spatiotemporal features of video data through 3D convolutional neural networks,and enhances the representation of receptive fields using multi-scale structure fusion.The network parameters of this model are small,and the anomalous behavior representation features are extracted on the whole video data,and the multi-dimensional Gaussian model is used for the abnormality discrimination.The network model is trained by video diverse behavior data,and does not require special training of abnormal behavior data,so that the model has certain unsupervised learning ability,which makes the algorithm have better versatility.It also reduces the difficulty of network training.3)The anomaly detection algorithm of this paper has the advantages of high precision and high speed.Under the general computing resources,it can meet the requirements of real-time monitoring.The experimental results on the public dataset show that the proposed three-dimensional convolutional neural network based behavior recognition and anomaly detection algorithms have achieved competitive results.In terms of behavior recognition,the proposed model is better than all traditional methods in recognition accuracy.Compared with the current algorithm using neural network,the accuracy of this paper is comparable to the current mainstream without the auxiliary classification algorithm.algorithm.In terms of anomaly detection,the algorithm of this paper has the best error rate index based on pixel standard.The error rate index based on video frame standard is optimal in the absence of training.