A Study On Object Tracking In Sequence Images

With the continuous development of computer technology and great improvement of computing power, using computer to realize the human visual capabilities has became one of the hottest research topics in the computer science field. Object tracking with image sequences which means getting object’s information such as shape. position, velocity, etc. through objects detection, locating and tracking has great significance to achieve the object behavior understanding and carry out further analysis and processing of vision-based applications.After years of research, the existing object tracking technologies still face many challenges, including overcoming variety of interference factors such as scene interference, illumination changes, pose changes and so on, which improves object tracking accuracy at the same time of ensuring that it meets real-time requirement. This dissertation thoroughly explored the key technology of object tracking by investigating the existing work:the multi-feature fusion, the precise object contour extraction, adaptive template updating, and proposed innovative solutions to deal with interference. The main contents are as follows:1. Multi-feature fusion mean-shift object tracking. Aimed at the defect of classical mean-shift object tracking algorithm which is susceptible to background interference due to using single color feature, an improved texture feature and color feature combination mean-shift tracking algorithm is presented. Feature joint similarity was introduced for the first time. Through a more comprehensive description by feature combination, the anti-interference capability and accuracy of object tracking was enhanced.2. A level set image segmentation method using prior information. Existing image segmentation methods are usually isolated, alone on the image segmentation. Moreover popular region-based level set image segmentation methods always assume that the image has two sub-regions and the segmentation results from maximizing statistical distance between the two sub-regions are often less than ideal. Aimed at these issues, a novel level set image segmentation method was presented to improve the accuracy of image segmentation.3. Mean shift object tracking based on active asymmetric kernel function. The symmetric kernel function used in the classical mean shift tracking algorithms contains many background pixels which would affect the tracking accuracy and stability, and the contour of the object in image sequence often changes. For this. a mean shift tracking algorithm based on active asymmetric kernel function was presented. During the tracking process, the asymmetric kernel function adaptively updated, and thus the tracking accuracy and reliability improved.In-depth analysis and adequate experiments show that the proposed methods are more robust and accurate than existing methods.