Crowd Abnormal Event Detection Method Based on 3D Pyramid Image Generation Network

Abstract

The present invention is based on a three-dimensional pyramid image generation network-based crowd abnormal event detection method, and relates to a method for identifying graphics, using a three-dimensional pyramid image generation network to generate an image, and comparing the generated image generated by the three-dimensional pyramid image generation network with the image to be detected The difference between the abnormal events of the crowd is detected, which overcomes the lack of motion information between image sequences, local features of different scales, and the small reconstruction error between normal images and abnormal images in the existing technology based on generated images for abnormal detection. Abnormal behavior flaws.

Description

technical field [0001] The technical solution of the present invention relates to a method for recognizing graphics, in particular to a method for detecting abnormal crowd events based on a three-dimensional pyramid image generation network. Background technique [0002] With the increasingly prominent public security issues and the popularization of video surveillance equipment, crowd anomaly event detection based on video surveillance can detect abnormalities in the crowd in time and avoid unnecessary losses, so it has important research significance in the field of public security. [0003] There are two main ways to detect abnormal crowd events: the traditional way and the way based on deep learning. The traditional method mainly extracts features from the aspects of optical flow and gradient, and then uses SVM for classification. In the traditional way of crowd abnormal event detection, due to the characteristics of its own algorithm, only some simple and basic feature...

AI Technical Summary

Problems solved by technology

[0004] In 2016, M Ravanbakhsh and M Nabi used the convolutional neural network (Convolutional Neural Network, referred to as CNN) analyzes the behavior of crowded crowds, and proposes to combine high-level semantic features extracted by CNN and optical flow feature modeling. However, due to the uncertainty of the types of abnormal events and the difficulty in expressing abnormal events, CNN with supervised learning obviously cannot play its own role. advantages and characteristics, and the introduction of the optical flow method to extract the movement features of the crowd still has problems such as being easily affected by light.

In 2017, SC Yong and HT Yong used Auto-Encoder (Auto-Encoder) and Long Short-Term Memory (Long Short-Term Memory, LSTM) to combine space and time in the paper "Abnormal Event Detection In Videos Using Spationtemporal Auto-Encoder" published by Springer information to detect abnormal crowd events, but the image restoration degree generated by the self-encoding network of unsupervised learning is low, and the key information in the image is easy to be lost, which affects the accuracy of the model

CN108280408A discloses a method for detecting crowd abnormal events based on hybrid tracking and generalized linear model, which has the defect of poor versatility in identifying pedestrian abnormal behavior by tracking pedestrian movement trajectories and extracting feature points of the tracking path

CN107729799A discloses a visual detection and analysis early warning system for crowd abnormal behavior based on deep convolutional neural network. In its algorithm, local features at the pixel level extracted by optical flow method are used. Pixel errors will directly affect the feature descriptor. Moreover, the optical flow method has a large amount of calculation, is easily affected by illumination, and only extracts the feature vector for each video frame, ignoring the defects of the correlation relationship between video frames and frames to a certain extent.

CN103258193B discloses a group abnormal behavior recognition method based on KOD energy features, which has the disadvantage of adaptively selecting the threshold value of the added group kinetic energy and directional potential energy, resulting in low adaptability of the algorithm in different scenarios

CN106778595A discloses a method for detecting abnormal behavior in a crowd based on a Gaussian mixture model. In this method, when a long-term static object in the background suddenly moves, false detection and missed detection are likely to occur in the foreground detection. The shadow suppression effect of moving objects is not good, the model modeling has a large amount of computation, and the algorithm has many steps and time consumption, and cannot guarantee the real-time performance of alarm information in video surveillance.

[0006] Although the existing technology has proposed various improved methods for crowd abnormal event detection by generating generated images that are closer to the original images, the quality of the generated images is still not high, the spatiotemporal information and local information of different scales in the image sequence are lost, and the normal images and The reconstruction error of abnormal images is small, the training time is long, the network training is unstable, and it is not conducive to distinguishing between normal images and abnormal images.

Images