The Research On Moving Object Tracking Technology In Image Sequences

The research on moving object tracking technology aims at detecting, identifying andtracking moving objects from video sequences, and furthermore understanding and analysisobject activities. Related techniques have been broadly applied in video monitoring, robottechnology, image retrieve image compression and so on. Therefore it is of great significance tostudy Moving object tracking technology. In this thesis, the moving object detection andtracking are considered. The main research achievements of thesis include:(1)In the case of Multi-Objects, the traditional algorithms often have the wrong targets andeven lost the objects, it also have the worse robustness and low real-time. According to thesituation, first of all, we propose the way which combined Mean Shift Model with SIFT featurematching algorithm. In each iteration of the Mean Shift model, we calculate the similarity of theMean Shift model and SIFT algorithm respectively according to the feature vectors of theoriginal and candidate targets, and then compare them. The algorithm with bigger similarity isused to locate the object position. The loop will be end when the position was accuratelylocated. In the above operation, the operation area was limited to the two times the sizes thanthe original, thereby reducing the amount of computation and improving the real-time. Secondly,the introduction of the α-β-γ filtering prediction algorithm is used to solve the target occlusionproblem and it also can predict the position of next image, the way can make us position theoperating area and improve the real-time.(2)The problems of object occlusion and the similar target interference in the multi-objectsaffected the quality of the tracking. According the problems, This paper presents the improvedthe camshaft algorithm which combined the camshaft and multi-features, and added the GM(1,1)algorithm, Firstly, in order to improved the probability distribution of the diagram, TheCamshift with the SIFT algorithm is not used the HSV space but the RGB space, because thespace contain more image information. The SIFT features are accounted to get the more accurately histogram. After further research, the paper added the edge features so that it canimprove the probability histogram. For the situation of there are similarity targets in the image,we combine the improved camshaft with texture features. The system can determine the realtarget area based on the texture similarity. Finally, the introduction of GM(1,1) predictionalgorithm make the occlusion problem solved and improved the real time. The experimentalresults show that the algorithm, proposed by the paper, have better tracking robustness andreal-time.