امید قهاری

Thesis: Real time moving targets detection and separating target from background in moving camera systems

Abstract
In general real-time detection and tracking of moving targets is a combination of image processing, automation and information science that will shape a new technology, which is able to quickly identify the object and its position in the image for the secondary purpose of tracking and recognition targets. Modern searches and surveillance operations conducted by UAV can be very complex and costly. These systems often used to search in the vast lands such as mountains, seas and deserts. In this thesis, the problem of moving target detection in aerial imagery has been studied. Since finding the moving regions are used widely and presented solutions accuracy is not good, this issue was an open problem and need to research and further investigation. Therefore, this thesis proposes a hybrid approach includes spatial and temporal information in order to deal with this problem. In this thesis, the temporal difference data used to determine targets approximated location and candidate regions that may be conclude moving objects. Then the saliency data on the candidate locations were used to obtain more precise information of moving targets. The data from the two modes combined to obtain location and exact specifications of moving targets. In temporal saliency detection section, images resized to 240*320 pixels to reduce additional computation. Because in this section need to find a coarse detection of moving objects locations. Also in the feature points detection section apply two constraints (number and distance) to keep accuracy and speed up feature point detection, description and matching sections. In this thesis, for having better information on spatial saliency, the three different types of saliency in the three fields is used to motion saliency information not duplicated and can complete each other. Finally, this algorithm is coded in C++ programming language and the images from VIVID dataset used to compare this method with other existing methods, and its performance is shown to be best. The results of this thesis show that the proposed method is capable to detect moving targets from aerial photographs in about 22 milliseconds, in a way that does not create fragmentation between the detected targets location and compared to similar methods of determining the coordinates and dimensions of various targets have achieved good accuracy.