Research On Robust Video Object Detection And Tracking Algorithms With Applications

Object detection and tracking in video stream is an important task in many computer vision applications such as augmented reality,intelligent monitoring,video retrieval,traffic detection,vehicle navigation,and object identification.Therefore object detection and tracking is a hot research field,whose theory and technology have been dramatically improved.Many interference factors bring in challenges for the visual tracking in real life,such as uncertainty of moving targets,changes in the observation characteristics of the target,and complexes in the environment.Although many researches have conducted in depth studies on target detection and tracking technologies in recent years,there are still many key problems unresolved.This dissertation focuses on methods for single target detection and tracking.In order to track a target under the specific environment,a tracking algorithm with theoretical guarantee is proposed and analyzed via simulations.The proposed algorithm is developed within a Bayesian modeling framework and the state inference is performed by employing the particle filter.Moreover this thesis also presents our studies on both dynamic state transition and observation models in the context of video object tracking,with the aim to reduce tracking error and improve tracking precision.Based on the above work,a novel tracking scheme that fuses multiple dynamic models and observation models is proposed.In this scheme,the dynamic model is represented by a combination of several basic dynamic models,each of which covers a different type of motion.The observation model also includes multiple basic observation models,which are related with a set of feature templates.Each basic observation model covers a specific appearance type of the object.Then a set of basic trackers are designed,each associating with a basic observation model and a basic dynamic models.All basic trackers are then integrated into the final tracking algorithm for our use.Experimental results show that our method tracks the object accurately with real-life video signals,in which both the appearance type and motion pattern of the target drastically change over time.