Adaptive Scene Modeling With Applicat Ions To Object Tracking In Video Surveillance

Nowadays video surveillance systems monitor the behavior of monitored objects in real time and play an increasingly important role in the tasks of managing and monitoring the city.It can be said to be ubiquitous.Video surveillance systems need to move in the direction of intelligence,with reducing as much manual intervention as possible or fully automated to accomplish the required tasks.Intelligent video surveillance technology mainly uses computer vision to achieve the required visual tasks,such as object positioning,object recognition,object tracking and other visual tasks in the scene.The process does not require any human intervention.Intelligent video surveillance technology automatically analyzes the sequence of video images recorded by the camera to understand and judge the behavior of the object in the sequence of images.And it can realize daily management needs,and can also make corresponding judgments and processes in time when abnormal situations occur.The intelligent video surveillance system is a challenging research direction and still attracts research from domestic and foreign scholars.When automatically analyzing image sequences,intelligent monitoring not only analyzes the characteristics of individual objects,but also analyzes the scene background and surrounding context of the object,that is,the scene information can provide an effective prior to the research of object activities.By studying the surveillance videos of a long time stored by the camera,the activity law of the object in the scene can be learned.Since the activity law of the object is affected by the scene background information,the scene information can be obtained by analyzing the object activity law.Scene information can be useful to assist in the study of activities for a specific object.In this paper,we adaptively model the scene,and fully exploit the scene information to assist the analysis of the object activity.We study the influence of the geometric characteristics of the monitoring scene on the scale change of object in the scene,and automatically learn the object activity law in the scene with the data-driven method.The state prediction of the object is achieved,and both research results are applied to the object tracking to achieve robust scale adaptive object tracking.There are two main results of this paper:1.The scale estimation of the object is achieved by analyzing the geometric characteristics in the monitoring scene and then is integrated into the tracking to achieve scale adaptive object tracking.In this paper,we propose to use the geometric context of the surveillance site as a strong clue for scale adaptation.With three reasonable assumptions on the video cameras and the surveillance sites,we deduce a simple geometric model for object scales.The parameters of this model are learned without any human intervention.Then we propose a generic method to integrate this model into some baseline trackers for robust scale adaptive object tracking.The scale estimation method proposed in this paper is based on scene geometry rather than object appearance,and is robust to many challenging factors(such as illumination changes,occlusion and fast motion,etc.).Experimental results on challenging surveillance videos indicate that our approach favorably improves the performance of single-scale baselines,and performs better or comparative to the state-of-the-art multi-scale trackers while significantly improve the speed.2.By analyzing the influence of the scene on the object activity,an object state prediction model based on the object motion model of the monitoring scene is proposed.With the data-driven method,the LSTM network is used to learn the data of object trajectories acquired by the object detection method without manual interference.The motion law and scale change law of the object in the scene can be learned at one time,and then the subsequent state of the object in the scene,including the object position and scale,is predicted.The experimental results show that the state prediction method proposed in this paper has a good effect on both position prediction and scale prediction.Then a general method for integrating the state prediction model into the tracker is proposed.Taking the existing tracker as an example,the robust scale adaptive object tracking is realized by combining with the proposed state prediction model.Experimental results on challenging surveillance video datasets show that the proposed method effectively improves the performance of the baseline tracker.And the method proposed in this paper learns and analyzes the data obtained from the object detection method without manual labeling,so it has a strong ability to scene migration.