Machine Vision Object Tracking Algorithm

Machine vision make the machine have the ability of observe, analyze andjudge like eyes by intelligent algorithm, and target tracking is one of the coretechnology of machine vision. Machine vision target tracking have a wide rangeof applications in both military and civilian fields, such as missile guidance,aircraft surveillance, intelligent transportation, medical imaging, etc., so thestudy of machine vision target tracking algorithm has a broad applicationprospect and practical value, has received attention from the academic circlesand business circles both at home and abroad.Machine vision target tracking algorithm research involves image processing,automatic control, machine learning and pattern recognition, and many otherdisciplines, with the development and mutual confluence of various disciplines,new ideas and new framework were certain to appear, and made remarkableachievements. But more and more high precision, speed, versatility andintelligence demand makes the study of machine vision target tracking algorithmis still faced with enormous challenges.Machine vision target tracking exist the following several main problems inthe real complex application environment: the target space rotation, scalechanges, fast moving and shade; Illumination change, similarity of backgroundand target, etc. How to overcome the above problems, and improve the speed,accuracy and robustness of target tracking algorithm is the important anddifficult of research. According to the advantages and disadvantages of theexisting algorithm, and analyzes of the differences and similarities of detection,tracking and recognition algorithm, this paper did the following innovation andwork in order to improve the target tracking algorithm performance:1.This paper proposes a quad-tree partition model of suitable for targettracking. The scheme to avoid the past blindly to extract the target feature, butwith different structural characteristics of tracking object, detailed segmentationto target with complex structure, rough segmentation to target with simplestructure, as well as the characteristics of the block with the method ofparameter estimation for description, both fully describes the target and greatly reduced the feature dimension, so as to accelerate the speed of target trackingalgorithm.2.This paper proposes a target tracking algorithm based on bag of words.Bag of words is a kind of abstract features learning methods of imitating humancognitive, the method in the field of machine vision detection and imageretrieval has achieved fruitful results, but due to the uniqueness of targettracking application, traditional classification by bag of word is difficult to beapplied in target tracking. A new feature and an online learning target trackingmethod is proposed in this paper, and the feasibility of the algorithm is verifiedby experiment, compare with the manual feature tracking algorithm the targettracking algorithm based on bag of words has better robustness.