GPU accelerated real / semi-real time object tracking

One of the most common methods used for feature detection is the point based feature extraction methods and the SIFT method is the most well-known. In this thesis, SIFT and SURF methods are used. Also the GPU-based SURF method are used to accelerate feature extraction. In this study firstly, SIFT, SURF and GPU based SURF algorithms are used for image matching and object tracking. Experiments show that, these methods give good results against image changes such as scaling, rotation, illumination, blur and affine transformation and can detect the object accurately. However, despite the stated success of these methods, the low success rate in object tracking applications is a problem. Against this problem, we present a novel object tracking framework for interest point based feature extracting algorithm such as SIFT, SURF and GPU-SURF methods. The proposed framework uses the feature extraction algorithm without making any changes. In the proposed system, it performs object tracking in three steps using the obtained features. These steps are outlier detection, object modeling, and object tracking. In the outlier detection step, after the keypoints are extracted by using a feature extraction algorithm, incorrect keypoint matches are detected by using the DBScan algorithm. In the second step, the object model is defined as a bounding box. The box model has six values and each of these values have its own Gaussian model. With the combination of DBScan and Gauss methods, object position can be accurately predicted even when the object feature points are few and inaccurate. Finally, the Gaussian smoothing process is performed for object tracking. The developed point based object tracking application works in harmony with SIFT and its variants. In addition, other point based feature extraction algorithms can be added to the object tracking framework with a short integration process. According to the results of experiments carried out in computer environment show that the proposed tracker improves the success rate of the object tracking significantly without affecting the execution time.