Research On Target Tracking And Trajectory Anomaly Detection

With the rapid growth of video surveillance data,the way based on artificial mode is not only time-consuming and laborious,but also inefficient.It cannot meet the actual needs of people.Therefore,intelligent video analysis is particularly important.A good intelligent video analysis system will bring great convenience to people's lives.As a hot topic in the field of computer vision,target tracking has been not only widely used in video surveillance,but also in some other fields such as human-computer interaction,intelligent transportation,computer vision,medical diagnosis,virtual reality,etc.Different target tracking algorithms have different advantages and disadvantages.Aiming at the situation that the traditional tracking algorithm is easy to cause the loss of target when the target object is partially occluded,the light changes suddenly or the target moves so fast,an improved target tracking algorithm based on color and texture features is proposed to improve the tracking accuracy in those situations.The main work of this thesis is as follows.(1)The thesis use three groups of targeted experiments to verify the Mean Shift algorithm in several specific scenarios has low tracking accuracy,in order to show my own algorithm can improve the tracking accuracy in those complex situations the next chapter.(2)Joined with the texture feature which is illumination invariant in feature selection,and in order to take the image spatial information into account,the blocked method we use also improves the accuracy of tracking.Aiming at the defect that Mean Shift algorithm is easy to loss target when the target is partial occluded,the method of neighborhood search is proposed.That is,when the similarity between the candidate target model and the target model is lower than the threshold value,we can see this as target being occluded.Then we extend 4 candidate regions of equal size near the candidate region,and calculate the similarity between the target region and the five candidate regions,choose the candidate region with the highest similarity as the final candidate region,and its center will be the position of iteration.Due to the expansion of the search area,it also has a certain improvement on the tracking loss because of the target moves so fast.(3)The thesis introduce several commonly used methods of trajectory distance measurement.Based on this,we analyze and compare the rationality of Euclidean distance and DTW distance.And propose our own distance measurement method.That is,the space distance metric using DTW distance metric,at the same time,we add the direction distance to it.And we introduce the clustering algorithm which is used in this paper in great detail.And this method is combined with the trajectory distance measurement method we proposed to cluster the trajectories and detect the abnormal trajectories.The validity of the method is proved by real experiments.