The dissertation at hand is concerned with the scientific solutions for the problems of visual object tracking and image-based railway quality control. In the tracking scenario, an object of interest is followed automatically. The type, size and, general features of the object are not known and the system does not have any knowledge of it. The object tracking schemes are studied based on three different sensors namely rectangular, omnidirectional and, 360-degree cameras. For the rectangular videos, two object detection-based methods are proposed. The first tracker uses a pre-trained object detector and does not update, whereas the second one is updated during the tracking. These systems compare five famous training based object detectors. The results demonstrate the trackers’ efficiency and show the preference of offline tracker to the online one. The polar cameras - namely omnidirectional and 360-degree sensors - provide videos
with a wider field of view than the conventional normal rectangular ones. Replacing conventional security cameras with 360-degree ones allows a significant reduction of hardware costs as well as software license and maintenance costs. There are many trackers based on conventional rectangular videos in the literature, whereas the number of polar object following systems, in comparison, is very limited and they are not yet matured. Most of the projects which are going to be discussed in this work are processing the 360-degree videos. Two proposed methods unwrap the polar videos using image rectification; then a modified version of Tracking Learning Detection tracker and another state-of-the-art detector are applied. To increase the speed of the process, a trapezoidal tracker is proposed to eliminate the rectification part. In another proposed scheme, a SURF based algorithm is used to improve performance. This tracker uses two learning-based modules for interesting points matching and challenges recognizing respectively.
The other proposed method combines a polar candidates generation method and color binary features to improve its accuracy and speed. The experiments show that the last method has the best accuracy and speed among the proposed methods and it outperforms the state-of-the-art polar trackers.
In the second part of the dissertation, a vision-based quality control method of concrete railroad sleepers is presented. This system captures an image sequence by a high resolution, fast and moving camera from railway top view and applies a proposed image-based crack detector to control the railway sleepers quality. This scheme first locates the sleepers within the images and then, generates crack candidates on the sleeper images and finally, detects and classifies the cracks and by applying a supervised classifier on the candidates. The classifier uses geometrical features to detect and classify the cracks on the concrete sleepers. The experimental results show that the crack detector successfully finds the cracks.