A Study On Human-Computer Intereaction Based Target Selection In Visual Tracking System

Object tracking algorithms have been well improved in recent decades and therefore are widely used in real world visual applications,like cellphones,drones,unmanned vehicles and surveillance systems.The practical application of object tracking research is faceing many unsolved issues,one of them is how to specify initialization information.In research stage,a tracker is initialized from the hand annotated image region in the first frame which is,however,not directly available in real world systems.Automatic object detection and human-computer interaction are both options to give the initialization information.Automatic object detection is not applicable in many visual systems.Model-based methods only detect objects whose prior models or training samples are given.General object detection methods,like moving object detection and salienct object detection,need certain object movement model or scence clutterness.Therefore,manually drawing object region is an indispensable method to initialize a tracking process in various visual tracking systems.Unfortunately,human-computer interaction based tracking initialization methods are not well investigated and no public human input dataset is available online.This dissertation is dedicated to both interaction dataset and interaction methods and to this end proposes several novel algorithms.The first public human-computer interaction dataset is established.We recruit several persons to collect two kinds of inputs under three different interaction constraints on various videos including different scences and target.More than 20,000 manual inputs have been collected,based on which we further model the human interaction operations.By analyzing and discussing major influencing factors,we divide the tested videos into two groups for more experiments.The proposed dataset reveals a clear gap in accuracy between manual inputs and the initial information of tracking algorithms.To solve this issue,we propose three interaction methods using different operationsa and exporting results with different accuracy.A single-click based interaction method is proposed.To generate a target region from single manual click,we propose a new object detection algorithm,which bases on object proposal cue and saliency cue.Object proposal is to find all potential regions containing arbitrary onjects and the generated regions tend to stay close to complete object contours.Based on this feature,the proposed general object detection algorithm is established and is independet to prior models by combining the proposals,saliency map and the manual input distribution given in the proposed dataset.Using a single manual click,this interaction method is simple and therefore is applicable to real-time visual systems.Experiments demonstrate that the proposed is effective to improve manual input accuracy and is successful to initialize tracking systems.A video stabilzation and bounding box drawing based interaction method is proposed.Traditional video stabilization methods remove camera jitters by estimating global motion.In synthesized videos,both target and scence move smoothly,but remaining object motion which decreases the interaction accuracy.To solve this issue,we propose a video stabilization algorithm removing both camera and target motion by estimating the target trajectory.In videos synthesized by the proposed algorithm,targets stay stationary on the interaction screen,which is benificial for more involuted human-computer interactive operations.Experiments demonstrate the proposed method can effectively remove both camera jitters and target motion,and on the stabilized videos,interaction operations can give results that are more accurate.This work enables more difficult interaction operations in more sophiscated environment.An online video segmentation based contour-aware interaction method is proposed.To support this work,we propose a new temporal superpixel segmentation algorithm,which proposes superpixel supporter group to deal with occlusions.Temporal superpixels propagate along image sequences,which means every superpixel has a corresponding region in another frame.Based on this feature,users can select a contour-aware target region by clicking corresponding superpixels in an arbitrary frame.The superpixel supporter group,togeteher with implicit shape model and generalized Hough voting,helps in finding occluded superpixels after occlusions.Experiments demonstrate that the proposed temporal superpixel algorithm is effective to increase superpixel trajectories and the proposed interaction method generates more accurate and flexible target regions.