Ship Visual Tracking For Key Inland Waterway Regions

Real-time ship tracking is of critical importance to inland waterway surveillance systems.As existing systems such as AIS and radar have distinct shortcomings with respect to coverage and investment-benefit ratio,visual techniques-based surveillance systems become the new trend.This thesis aims to provide effective and efficient solutions to ship visual tracking problem in inland waterway CCTV scenes.To the best of our knowledge,this thesis is one of the earliest work that aims to attack the challenge of inland waterway ship tracking from a pure visual perspective.The main work can be summarized as follows:(1)First,a thorough overview of more than 50 existing state-of-the-art randomprojection based visual trackers is conducted from different perspectives,including random measurement matrix,low-dimensional features,feature selection strategy and so forth.Second,extensive theoretical and experimental analyses are outlined to provide readers a good understanding of the strengths and weaknesses of different trackers,especially in the domain of inland waterway ship tracking.These observations motivates us and other researchers to provide better solutions to perform robust inland waterway ship tracking.(2)Within the Bayesian state estimation framework,a novel random-projection with low-rank regulation ship tracking algorithm is proposed.First,our state transition model can predict the state variation well by differentiating the state between consecutive frames.Second,orthogonal experiment design strategy is employed to repermutate and evaluate the particles.Third,contour and color features are effectively fused to better describe the target ship and a new random measurement matrix is proposed to account for different scale initialization.Finally,our observation model is constructed via random projection with low-rank regulation,through which the strong assumption on data's prior distribution is alleviated.(3)Besides,a novel tracking and detection based inland waterway ship tracking algorithm is further proposed.The tracker and detector runs simultaneously and independently,and the final tracking result is integrated with the outputs of these two modules.The search regions for the detection module are adaptively determined by the output of the tracker module,which acts as an indicator of the appearance change of the target ship.The proposed random-projection with low-rank regulation detector demonstrates superior performance,which further facilitates the cascade detectors to separate the target ship from the background.(4)To the best of our knowledge,there does not exist specific benchmark and online evaluation system that are publicly available in the inland waterway ship tracking community.To address this issue,a comprehensive and representative inland waterway ship tracking benchmark is developed with data collected from local maritime bureau.Furthermore,an online evaluation system is designed to hopefully provide researchers uniform basis to evaluate and analyze different algorithms.