Anomaly detection and localization method for unsupervised crowd based on recurrent neural network modeling

Abstract

The invention discloses an unsupervised crowd anomaly monitoring and positioning method based on recursive neural network modeling. The method combines the time series characteristics of the surveillance video with the long-term dependence of the recurrent neural network, and divides the video scene into a grid , model each grid separately, selectively use optical flow statistical features to learn unsupervised crowd sequence dynamics under normal conditions, and use the Hessian‑Free Optimization method to train the model, and finally will contain abnormal data Load the model, and monitor and locate the crowd scene at time t+1 by measuring the distance between time t+1 and the statistical histogram of optical flow at time t. This method realizes the global modeling in time and space, well preserves the connection between the dynamic sequence features of the crowd at each moment, and simplifies the complexity of the model, reduces the difficulty of RNN training, and monitors and locates abnormal events It has better real-time and accuracy.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to an unsupervised crowd abnormality monitoring and positioning method based on recursive neural network modeling. Background technique [0002] Video surveillance detection technology is an image application technology that combines video technology and modern communication technology. At the same time, the detection of abnormal crowd events has also aroused the interest of more and more researchers around the world. [0003] Rao, S. et al[1] developed a stochastic model in 2003 to describe the behavior of normal people. When a new video is sent, the motion trajectory of the person is extracted to test whether there is any abnormality. In 2004, Shobhit Saxena[2] and others proposed a multi-frame feature point detection and tracking algorithm based on KLT tracking to realize crowd event modeling in specific crowd situations. Their proposed extended scene recognition engin...

AI Technical Summary

Problems solved by technology

However, in the actual research process, whether it is in the initial extraction of crowd dynamic features or in the later motion analysis, there are a series of problems that lead to a high false alarm rate in intelligent video detection, such as: different, the background is complex, and the definitions of basic events given are inconsistent; there are many types of abnormal events, which rely too much on manual definitions. For a large number of video surveillance, the workload of manual labeling is also increasing, and it is easy to ignore some specific Non-violent anomalous incidents such as onlookers or stepping on grass

So far, the research on the two basic problem-solving methods for the detection of abnormal crowd events is still in the immature stage, and further study, research, and improvement are needed.

Images