Detection Based Online Multiple Object Tracking

Object tracking, which aims to estimate object trajectories in video scenes, haslong been an active research topic in computer vision. Recently, with the fastdevelopment of object detection techniques, detection based multiple object tracking isbecoming a very important class of tracking problem and has many practicalapplications as wide as visual surveillance, traffic monitoring, and content-based videoretrieval. It is also a quite difficult problem due to abrupt motion of objects and complexinteractions among them, which requires designing many novel and sophisticatedmodels and algorithms. Thus it is also very important in academic research.This thesis focuses on how to online track multiple objects simultaneously based ondetection, which is an even more challenging problem because only the imageinformation until current processing frame can be used to infer for the most possibleresults. By means of systematic in-depth analysis of the tracking problem, it has beenfound that the observation model and the tracking strategy are two key components inthis problem. From this perspective, several new online multiple object trackingalgorithms on many typical kinds of video scenes, including surveillance videos,consumer videos and sports videos, have been proposed and successfully applied intosome middle and high level applications. The main work of this thesis includes:Firstly, a two-stage online multiple object tracking algorithm by local trackletsfiltering and global tracklets association is proposed. It designs a multiple lifespanbackground subtractor and a multi-part multi-view human detector, both of which areadapted into a local tracking procedure using a particle filter with selection and a globalassociation tracking procedure within a temporal sliding window. Good results onoccluded objects tracking is obtained through this method.Secondly, a multiple object tracking algorithm based on online discriminativelearning is proposed. Through online learned discriminative features from interestpoints and color patches, the ability of the tracking system to deal with more seriousocclusions and distinguish between every two different objects are enhanced, whichfurther improves the tracking results of occluded objects.Thirdly, an approach based on progressive observation modeling and dual-modetwo-way Bayesian inference is proposed for multiple player tracking in sports videos. The progressive observation modeling process divides an initially difficult problem intosome solvable sub-problems and tackles them step-by-step to collect robust and reliableobservation information sequentially. These observations are directed at a unified singleobject and multiple objects tracking procedure by forward filtering and backwardsmoothing. The whole algorithm provides a very good solution for the problem oftracking multiple players with abrupt motions and complex interactions.Last but not least, two new algorithms, which respectively address the humansegmentation and crowd counting problem, are proposed based on object trackingresults and techniques. The experiments demonstrate that the processing results of thesetwo middle-and-high level vision analysis problems can be significantly improved withthe help of object tracking.