Research On Detection Technology Of Violence In The Background Of Monitoring

With the rapid development of human society,incidents in today's society have become more and more complex,and there may be more and more sudden abnormal events.This requires intelligent video surveillance to play a greater role in security.However,the vast majority of video surveillance now only serves the ability to preview and store surveillance video in real time,and it cannot intelligently analyze what is happening in the current scene.Many monitored videos need to send full-time personnel to watch the surveillance video and judge the events and behaviors in the scene,wasting human and financial resources.The advantage of deep learning based violence identification algorithm is that the recognition rate is relatively high.However,there are many limitations in the implementation of such algorithms.First,deep learning algorithms require a large number of samples to adjust the model during training.Some data sets and actual conditions do not provide enough samples for training.The recognition rate of the model obtained by the training of a small number of samples is relatively low,and from the production of the data set,a large number of samples need to be manually labeled,and the time and labor costs are high.Looking at the number of samples contained in an existing data set,it is not suitable to use deep learning algorithms to test the accuracy.Therefore,the paper chooses the algorithm based on machine learning to study.This article mainly focuses on the short video common data set of 50-80 frames and the video collected in the actual monitoring scenario as the research object,and proposes a new detection algorithm for violence behavior.The proposed algorithm performs frame difference and binarization calculations on the images of two adjacent frames,and then describes the image of the binarized motion blob attributes.Three features of the moving blob's area,compactness,and distance between centroids were extracted,and the optical flow energy histogram was added.These features were laterally stitched into descriptors,and these descriptors were processed as feature vectors.Send to random forest classifier for training and identification.Experiments were conducted on public data sets Hockey,Violent flow video sets,and self-made data sets based on indoor and outdoor surveillance video.The results show that the algorithm can not only achieve better results in public data sets,but also Good results in video surveillance scenarios.