Multi-Scale Feature Analysis and Object Tracking under Complex Scene
|School||Dalian University of Technology|
|Course||Applied Computer Technology|
|Keywords||Multi-Scale Scale Space Bayesian Theory ROC AUC|
There is an urgent need for efficient computer searching and tracking approaches in the natural environment int the field of modern medicine, aerospace, natural surveillance and military applications. These tasks typically have a high complexity and are demanding on the response speed due to the complexity of the natural complexity. With respect to simplicity and regularity of the artificial scenes int the laboratory, natural scenes in reality appear much more complicated. Under natural scenes, the background is rarely constant static and generally dramatically changeable and it is difficult to find variation rules, which obviously increases the complexity of the scene. In some complex scenes, several targets with similar structure, texture, gray distribution, edge and so on may exist in the same scene simultaneously, which makes targets identification very difficult. In addition, the illumination in the natural scene may change over the time because of the weather and some changes such as translation, rotation and random structural adjustment about the target itself may also happen, which all make searching and tracking of targets more difficult. Therefore, the study under complex scenes on target detection and tracking algorithms is of great significance in both theoretical and apllied aspects.On the one hand, against increasing scene complexity with changeable scales, this paper focus on multi-scale analysis of complex conditions and target tracking based on multi-scale features. The feature presentation generally differs in different scale conditions. The scale is likely to change when target moving relative to the camera. Therefore, multi-scale feature analysis based on automatic scale selection is particularly important. Meanwhile, when the target moving relative to the camera, the target size generally changes over time and the target tracking based on automatic scale selection can effectively resist against the environment complexity and noise. In this paper, a number of feature analysis and target tracking algorithms are presented and achieve good results in theory analyses and experimental comparisons.On the other hand, a parallel analysis framework with Bayesian theory on the background and the target is presented, which reduces the complexity of the tracking of the target under natural scenes. In addition, this paper reviews and sums up the feature integration theory as an effective complementation about analysis and synthesis of multi-scale theory, which increases completeness.