Research On Abnormal Behavior Recognition Algorithms Based On Deep Learning

With the rapid development of computer vision and in-depth learning technology,in the field of public security,a series of intelligent algorithms based on in-depth learning have a broad application market and great research value.In view of the huge scale of monitoring data,the existence of redundant and invalid data,and the lack of high-quality annotated data,how to use.in-depth learning to identify abnormal behavior efficiently and accurately is an important research topic.This paper is mainly divided into two parts to carry out the research of abnormal behavior recognition.(1)For retraining Inception-v3 algorithm,it takes too long and the recognition rate is not high,and it is prone to fit in the absence of high-quality labeled data.This paper presents a migration learning recognition algorithm based on pre-training Inception-v3.In the feature extraction.stagethe pre-trained Inception-v3 algorithm is used to extract features by migration learning.In the prediction stage,three-layer feed-forward neural network,k-nearest neighbor algorithm,logical regression,random forest,GBDT,LightGBM,SVM and two-layer feed-forward neural network are established for behavior prediction,and some prediction results are fused with Average algorithm.Experiments show that compared with the retraining Inception-v3 alg.orithm,the pre-training Inception-v3 has higher recognition rate,and is not easy to fit.On the basis of migration features,the prediction effect of the proposed feedforward neural network algorithm is better than that of the traditional machine learning algorithm;the recognition rate of abnormal behavior recognition is further improved by the algorithm fusion.(2)To solve the problem that the feature expression ability extracted by single spatial flow convolution neural network is not strong and the accuracy of anomalous behavior recognition is not high.In this paper,a dual-stream CNN+LSTM hybrid algorithm is proposed Spatial flow network is used to process RGB images,and temporal flow network is used to process optical flow images.Appearance features and motion features are extracted respectively to enhance the ability of feature expression.At the same time,LSTM is introduced to establish abnormal behavior recognition and classification algorithm.Compared with ordinary neural network,LSTM introduces input gate,forgetting gate and output gate to avoid gradient explosion and gradient dispersion to a certain extent,and can capture temporal information for classification.The experimental results show that the recognition rate of the two-stream CNN+LSTM hybrid algorithm is higher than that of the single spatial flow neural network algorithm.