Detection And Analysis Of Video-based Pedestrian Trajectory Extraction And Its Abnormal Behavior

At present,the scale of video data is increasing.Manual video analysis not only consumes a lot of manpower,but also increases the security cost,which can not meet the current development needs.It is very important to analyze video intelligently.Anomaly detection is the most important research direction of video analysis technology.Anomaly detection refers to the analysis of video content and the prediction of possible abnormal events or behaviors.The motion track of the object in the video information is rich in a lot of space-time information,which is of great value to infer the action trend and characteristics of the specific object.However,the main methods used to analyze the trajectory are relatively complex,and most of the algorithms are still in the theoretical research stage,with few practical applications.This research mainly discusses the routine program of detecting abnormal behavior based on trajectory,puts forward a very reliable way to obtain abnormal information from the target trajectory,and establishes a model and system that can effectively detect abnormal phenomena.This system includes many steps,such as detecting moving target,tracking target in real time,extracting pedestrian track and its characteristics,learning and detecting abnormal behavior.Firstly,the improved and updated hybrid Gaussian model is used to detect the foreground,import the reliability,prominent Gaussian distribution and variable learning rate,so as to accelerate the process of building the background through learning.Considering that the shadow used to detect the foreground will make the shape of the object wide and irregular,it will affect the target tracking.In this study,we use the basic method of maximum chromaticity difference in color space dimension to remove the shadow,obtain more accurate shape of the target object,and effectively reduce the probability of hiding the foreground area.Then,the subject of stable tracking target is studied.The two methods of average drift and matching foreground position are combined.According to the possible situation during tracking,the target state is divided into several categories.The wired automata is used to manage the target state,which changes and transfers the state of the moving object in combination with the occurrence of conditions.After the target formally enters the research field of vision,it completely leaves the field of vision The whole process of encirclement will be recorded.In addition,timers are designed for forbidden objects in the scene interval,and the dwell time of specific objects is detected.Finally,the acceleration,speed and length of each track are extracted,and the feature vector with fixed length is established.The self-organizing feature network is established to train and learn the feature vector.The feature of normal track points in the scene range is obtained.The topological structure is established.The feature values of target track points and normal track points are compared and detected,and the "abnormal" judgment is made The trace with more abnormal trace points is determined as abnormal trace.The method proposed in this paper can quickly detect abnormal trajectory points,and then real-time early warning of suspicious pedestrian trajectories.Experiments on a video taken in an open-air parking area show that this method has a good detection rate for abnormal trajectories.