Research On Technology Of Human Abnormal Behavior Recognition Based On Video

With the development of science and technology and the advent of the information age,pedestrian safety has received more and more attention as a part of public safety.It has important theoretical research value and social significance for pedestrian's video-based abnormal behavior detection.The current abnormal behavior detection methods have the following problems: First,when there are multiple detection targets or the detection environment is complicated and changeable,most of the moving target detecting and tracking algorithms have a large amount of calculation,a long calculation time,and Low stickiness.t the same time,it is difficult to determine the affiliation between each extracted feature value and a plurality of detection targets.The second is that it is difficult to define abnormal behaviors,it is difficult to extract their detection features,and it is difficult to distinguish them from similar behaviors.For example,the abnormal behavior of falling is easily confused with rapid crouching.1)Aiming at the problem of low robustness and time-consuming detection of multiple moving targets,the OpenPose algorithm will be applied.The VGG network is used as the overall framework,and two pre-trained network branches are used to return the confidence map of key point positions and the associated vector field between each key point.Perform multiple successive iterations on each branch to refine the overall pre-training steps of the network.After each round of iteration,a loss function needs to be calculated,and the keypoint position confidence map and associated vector field obtained from the previous round of iteration are fused with the original input as the input of the next iteration.In this way,the pose features of the detection target are obtained,and further analysis and processing are performed.2)Most of the researches on abnormal behavior recognition use self-built datasets,there is no more common and open behavior recognition datasets,and there is a lack of research status of horizontal contrast conditions.Before performing abnormal behavior recognition,the detection algorithm used in the mainstream behavior recognition dataset is verified by experiments and comparisons,and the confusion matrix and recognition accuracy rate are used as evaluation parameters for the result analysis.3)Aiming at the problem that it is difficult to define and distinguish between abnormal behaviors,several actions are set as abnormal behaviors,including: crossing the line,regional invasion,and fall.The mathematical and statistical methods are combined to detect the dynamic and static characteristics of the target and apply Part of the affinity field matching algorithm completes the pose feature matching.The problem that it is difficult to determine the membership of a plurality of detection targets is classified as an NP-hard problem.The extracted feature points are randomly matched,and the optimal weight of each discrete candidate key point pair is calculated to determine the subordinate relationship between each key point and the detection target.This paper proposes a modified dual-stream CNN algorithm based on skeletal joint point information and motion posture information.After verifying the accuracy of the algorithm on the UTKinect-Action3 D public behavior recognition dataset,the fall behavior recognition based on the self-built dataset was performed.Finally,the three abnormal behaviors of fall,line crossing and area intrusion were recognized.