Research Of Scalable Visual Tracking And Related Applications

With the rapid development of digital multimedia technology,the total amount of video data such as user videos and surveillance videos has been explosively increasing,which poses great challenge to Internet inspection and security monitoring.Therefore,it is urgent for researchers to make a remarkable progress in video content analysis systems and related techniques,e.g.visual tracking technology.Serving as a crucial component of video-related applications,object tracking has been a hot spot in recent years.The difficulty of the visual tracking technology lies in two ways.1)There are plenty types of noisy data exist in videos,such as illumination changes,partial occlusion,background clutter,camera motion and ever-changing scales,and a good tracker should perform well under all of these circumstances.2)The tracking algorithms should run in real time when detecting the target between two frames,otherwise the playback speed will be dragged slow.It can thus be seen that a qualified tracker should guarantee stability,robustness and efficiency in its “train-track-update” procedure.In this paper,we design and develop a scalable tracking algorithm based on kernelized correlation filters.Our tracker makes full use of the circulant structure in the dense sampling strategy to realize scalable tracking in one filter.It implements multiple edges tracking or part-based tracking by proposing a Gaussian training output matrix with four peak values.Unlike existing scalable trackers which perform the correlation filtering operation many times or extract many candidate windows in multiple scales,our tracker fixes the scale problem under the original MOSSE/KCF single-filtering framework with super real-time tracking speed of 166 FPS.After the filtering operation,the scalable tracking algorithm will get a four-peak-value response matrix,which is more complex than the original one.Instead of locating the target by a single peak value as the original single correlation filter methods do,we propose to use a weighted Bayesian inference framework to deduce the location and size of the bounding box from the response matrix.It adaptively weights each part response both spatially and temporally,and then makes a biased combination of all confidence maps based on the response weights.It is common that some part result may be influenced by partial occlusion or illumination changes,and in this way the result of a reliable part can be considered more.As shown in our tracking experiments,our method performs more effectively and robustly than the baseline method,namely the KCF tracker.At the end of this paper we design and develop a prototype system for content analysis in surveillance videos.The system is mainly divided into three parts: 1)an object detection module based on Faster R-CNN which detects interesting targets in the input video;2)the proposed scalable single-filtering tracking algorithm which keeps tracking each target;3)an abnormal behaviors recognition module which analyzes the targets' behaviors in some specific scene.The system assists its users by giving an alarm when a dangerous act is detected,saving them from watching the security monitors all day.