Video-based fighting detection method

Abstract

The invention discloses a fighting detection method based on a video. The method comprises the steps of based on a target detection method, effectively detecting a human body target in a video ; secondly, extracting human body skeleton key point information, including continuous multi-frame skeleton key point 2D coordinate information of each human body; constructing a skeleton sequence, and constructing a space-time convolution graph on the skeleton sequence; and inputting the space-time convolution graph into a trained multilayer space-time convolution network (st-gcn) to perform action recognition. The method can be used to accurately identify fighting and other action behaviors, can be widely applied to stations, airports, supermarkets, commercial streets, playgrounds and other important public places, and realizes real-time early warning.

Description

technical field [0001] The invention relates to the technical field of video intelligent monitoring, in particular to a method for detecting fights based on video. Background technique [0002] Fighting and fighting are violations of law and discipline, which have a negative impact on social stability and people's lives. For fighting and fighting incidents, at present, it is mainly to notify the police after the event, and it is difficult to get the police in the first place. come difficult. With the comprehensive application and improvement of the Skynet project, the fight detection technology based on video analysis has gradually developed. [0003] Currently adopted technology 1: use the two-dimensional image registration method to calculate the pixel motion vectors of adjacent frames of the video, perform statistical analysis on the motion vectors of the human body or convert them into energy forms, and extract pixels with chaotic motion directions and fast motion , th...

AI Technical Summary

Problems solved by technology

Patents based on this type of technology include: a detection method and device for violent crowd movement, patent number: 200910242555.6 and a fight detection system for joint analysis of audio and video, patent number: 200920291779.1, the defect of this type of technology: the movement of adjacent frames of the video Vector calculations all use the differential optical flow method based on gray-scale features to obtain the pixel position offset on a two-dimensional image, but the optical flow method based on gray-scale features has many defects that limit the promotion of this type of technology, for example, it is affected by The change of the lighting environment has a great influence, and the lighting change of the scene will cause the calculation error of the real motion field of the target, and the motion field when the target is occluded cannot be calculated; because the distance between the target and the camera is different in the 3D scene, the motion vector reflected on the 2D plane is different, Therefore, the traditional method is not robust enough for the determination of violent motion;

[0004] Technology 2: Use the method of combining stereo vision and sports field extraction to reduce and filter optical flow calculation errors caused by illumination changes, target movement confusion and interlaced occlusion, and improve the robustness of human fighting behavior detection in complex environments. Based on this technology The patents are: a fight detection method based on stereo vision sports field analysis, patent number: 201210304084.9, the defect of this type of technology: the method of combining stereo vision and sports field extraction, in the case of complex scenes, cannot detect human objects The location is detected accurately, and there are large false positives for strenuous movements like running

Images