Video data anomaly identification method and device based on deep video event completion

Abstract

The present application relates to a video data anomaly identification method and device based on deep video event completion. The method includes: obtaining the area where the foreground object is located in the current video frame, obtaining the image data of each video frame in the foreground object area from the preset length video frame sequence where the current video frame is located, and generating a video event according to the obtained image data sequence Data, extract image data from video event data to obtain incomplete video event data, input the preset deep video event completion model to perform appearance completion and motion completion, according to the error between the completion result and the extracted image data Identify anomalies in video event data. The above method uses a deep video event completion model based on a deep neural network, combines appearance completion and motion completion to complete incomplete video events, fully taps high-level semantic features in video event data, and utilizes the context in the video. information, effectively improving the performance of identifying abnormal video events.

Description

technical field [0001] The present application relates to the technical field of pattern recognition and video security monitoring, in particular to a video data abnormality recognition method and device based on deep video event completion. Background technique [0002] Identifying abnormal events from surveillance video data is the core task of intelligent security. However, due to the characteristics of abstraction (fuzzy definition), novelty (usually brand new and unrecorded situations), and low frequency (the frequency of occurrence is generally extremely low, making data collection difficult), the most commonly used scheme is single-category Classification, training a normal event model using data containing only normal video events, and distinguishing video events that cannot be described by the normal event model as abnormal events. [0003] Current technical solutions lack the ability to extract high-quality feature representations from video events. The core prob...

AI Technical Summary

Problems solved by technology

However, this paradigm only uses the prediction effect of a single video frame for abnormal recognition, and abnormal events in videos often last for multiple frames, and the context information in video frames other than the predicted frame cannot be fully utilized to identify abnormalities; and, When the video frame is used as the basic unit of processing, it often contains different scales of the foreground and a large number of irrelevant backgrounds at the same time, which will interfere with the deep network training and video event feature learning process, and reduce the network's ability to focus on video events.

Images