Research On Abnormal Behavior Detection Based On Video

The aggravation of the aging population in China has brought great challenges to the areas of guardianship and security.Due to the popularity and cheapness of video monitoring and the advantages of video visualization and easy storage,it quickly becomes an important means to replace manual management.How to efficiently process the obtained video data,so that it can meet the accuracy requirements of human behavior detection in specific situations,and at the same time take into account the efficiency and try to meet the real-time requirements,has been the goal of algorithm researchers continuously.At present,most video processing tasks are based on specific situations,and the types of behavior detection and recognition are relatively fixed and single.When the traditional method uses the prior knowledge and intuitive feeling to design the characteristics strongly related to the task,it will be much better than the deep learning method in efficiency and accuracy.In view of this,this paper will focus on video abnormal behavior detection based on traditional methods.The work of this paper mainly includes the following parts:(1)Considering the time-consuming problem of optical flow calculation in the traditional video behavior recognition algorithm,this paper adopts FlowNet2.0 network based on deep learning to improve the optical flow calculation in the dense trajectory algorithm.While maintaining the optical flow calculation accuracy,the optical flow calculation speed is significantly improved;(2)Considering that the dense trajectory algorithm only describes human behavior from the time dimension,but lacks the description of space dimension,skeleton joint node feature is added in the paper,which effectively makes up for the deficiency of dense trajectory algorithm in the description of space dimension,and the performance of the algorithm is significantly improved;(3)In view of the problem that the importance of different features should be considered during feature fusion,the paper adopts the method of weighted fusion,and USES the ratio of trace of the inter-class divergence matrix and intra-class divergence matrix as the measurement standard,so as to assign more weight to the features of higher importance in the stitching,for which the performance of the classification is significantly improved;(4)Considering that the simple binary classification model has a poor migration ability when the abnormal behaviors defined are changed,the multi-label behavior recognition algorithm is applied to the detection of abnormal behaviors in this paper,which enhances the usability of the model when the abnormal behaviors defined are changed;(5)The experimental platform was built,and five public video data sets such as UCF-101,HMDB51 and KTH were divided and the model was trained by using the set aside method and cross-validation.Then,the predictive performance of the model was evaluated by t he evaluation indexes such as accuracy rate,recall rate and F1 score,and the classification accuracy is significantly improved when the speed of optical flow calculation is increased by nearly 7 times and the speed of the whole behavior recognition process is increased by nearly 6 times,proving the effectiveness of the model improvement.