Research On Object Tracking Based On SIFT Feature And Mean Shift

Video object tracking is an important research subject in the prospect of computervision technique, and has got extensive applications in such fields as human-computerinteraction, video monitoring, intelligent buildings, and military area. Therefore, it isvery significance to research it. Mean Shift algorithm is one of the algorithms whichhave got the most extensive applications. For Mean Shift algorithm can not track fastmoving object, this thesis takes the advantage of SIFT (Scale Invariant FeatureTransform) features, and proposes a new object tracking algorithm based on SIFTfeature and Mean Shift. Besides, the tracking performance of the new algorithm isevaluated by several experiments. The main contents of this thesis are as follows:Firstly, Mean Shift algorithm is introduced and analyzed, and some experimentsare taken. In this algorithm, the grey or color statistical characteristics of the objectimage is as the target model which can represent the object well. The experimentalresult indicates the performance of tracking fast moving object of this algorithm is notgood.Then, the theory and process of SIFT feature extraction, the standard and priorityk-d tree search algorithms are respectively introduced. Based on the existing trackingalgorithm using SIFT feature, an improved algorithm is designed. Some experimentsabout SIFT feature extraction and matching, the existing tracking algorithm using SIFTfeature and its improved algorithm are taken. The experimental result shows that, SIFTfeature is invariant to image rotation, and when the grey or color distribution of imagesis intensive, the feature number is a little, and the improved algorithm can tack fastmoving object. The result also shows there is a defect of these tracking algorithms usingSIFT feature. The defect is that when the feature number is continuously a little or zero,these tracking algorithms fail.At last, in this thesis, the advantages of SIFT feature and Mean Shift algorithm arecomplemented, in order that a new algorithm based on SIFT feature and Mean Shiftalgorithm is designed. To evaluate the performance of the new algorithm, someexperiments are taken. The target model can represent the object well in Mean Shift, butthis algorithm can not track fast moving object. The improved algorithm using SIFTfeature can track fast moving object, but when the feature number is continuously alittle or zero, this algorithm fails. In the new algorithm, the grey or color statistical characteristics of the object image is as the target model, and the object location ispredicted by SIFT features, and Mean Shift algorithm starts to track from the predictionlocation. The tracking comparison results of Mean Shift algorithm, the improvedalgorithm using SIFT feature, the combined algorithm of Kalman filter and Mean Shift,and the new algorithm show that the performance of the new algorithm is better than theothers.