Research On Human Abnormal Behavior Detection Based On Deep Learning

With the gradual improvement of people's safety awareness,the recognition technology of abnormal human behavior based on video surveillance has been widely concerned by the whole society.However,it is difficult to identify the abnormal behavior accurately using the existing abnormal behavior recognition methods with the high computational complexity and poor model generalization.Therefore,the study of human abnormal behavior recognition method in video is of great significance.In order to make full use of the effective information in the video and improve the recognition rate of the abnormal human behavior in complex scenes,an improved target detection method is proposed based on the mixed Gaussian model by studying the commonly used target detection method.The background model is initialized by the median method.A new weight update rule is established,a clear outline of foreground moving object is detectedand the key moving region is further extracted by combining the inter-frame difference method with foreground moving targets detected.The Farneback dense optical flow algorithm is used to calculate the optical flow value of the key region to obtain the spatio-temporal information.By analyzing the behavior classification model based on the deep learning,a CNN-LSTM hybrid two-stream model based on Dropout mechanism is established with the CNN and LSTM network combined.With the original image and the superimposed optical flow image of the key moving region of the input video sequence imported,the dynamic and static features and time series information of space-time information are learned.The weighted fusion method is used to weight the Softmax output of the two networks to obtain the final classification result.On the basis of behavior classification,the recognition of abnormal human behavior in video is realized,and the improved two-stream model is tested and analyzed by using test samples.The test results show that the improved two-stream network model achieves the accuracy of 91.2%on the behavior classification of the test set,and the recognition rate of abnormal behavior is 92%.Compared with the three models in this paper,the improved model improves by 6%,8.3%and 3.4%respectively indicating that the improved model has better recognition effect and provides a theoretical reference for human abnormal behavior recognition research.