| Summary: | 碩士 === 亞東技術學院 === 資訊與通訊工程研究所 === 101 === Application of object tracking has always been an important issue in computer vision or image processing applications. In the early stages, object tracking had been applied to air traffic control. Recently, it has often been applied to with security monitoring related fields. There are various types of methods for object tracking. Generally, these methods can be divided into time domain methods and space domain methods. In a time domain system, the target must be able to show time differences. In other words, the target has to move so that judgments can be made. In a space domain system, judgments are made based on image characteristics of the targets. And usually judging methods based on characteristic information are more complex and diversified. However, if only a time domain method is applied, the only thing that can be confirmed is that whether the target is moving or not. It is difficult to find out if this target is the interest one. On the other hand, particle filter is a good solution for target tracking but is only applicable when the scene is known and possible locations of the target are preset. However, adding the adaboost algorithm helps to solve this issue. Therefore, this study proposed a combined structure with adaboost detection and particle filtering method to resolve the problems mentioned above for the pedestrian tracking problems.
Considering this problem, a hybrid structure combining adaboost classifier and particle filter is proposed to automatically detect and track the pedestrian targets in this paper. The adaboost detection process is adopted first to target candidate objects, and then the particle filter is applied for confirming and tracking of targets. Experiment results show that via the proposed method, the drawback of the current particle filters which requires specifying an object to be tracked in advance can be overcome, while performing good also in cases of target missing, occlusion, and identifying the previously appeared objects.
Like other current methods, the issue of bad performances in detection and tracking in a complex environment still exists with the object tracking method proposed by this study. In the future, we will add a preprocessing step of image enhancement before target detection and increase the number of negative samples in the sample for cascade training, to solve the issue above, so this system can be applied more widely with efficiency.
|
|---|
