Image Sequence Analysis for Object Detection and Segmentation
Open Access
- Author:
- Gandhi, Tarak L.
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- December 14, 1999
- Committee Members:
- Rangachar Kasturi, Committee Chair/Co-Chair
Octavia I Camps, Committee Member
Lee David Coraor, Committee Member
Rajeev Sharma, Committee Member
David Jonathan Miller, Committee Member - Keywords:
- Image Motion Analysis
Target Detection
Computer Vision
Image Processing - Abstract:
- A sequence of images contains more information than a single image. Due to this reason, image sequence analysis has been used in computer vision for quite some time. In particular, a sequence of images is useful for object detection when the camera moves relative to the object. Due to the relative motion, objects at different distances from the camera have different image motion. Using this property, one can obtain information on structure of the 3-D scene as well as the relative motion between the camera and the scene. Furthermore, the individual images are corrupted by camera noise. Use of a sequence of images enables suppression of this noise for reliable detection of low contrast objects. In this research, image sequence analysis is used for detection of objects in 3-D space as well as on planar surfaces. The work would be useful for detecting obstacles in the flight path of an aircraft. The research also explores the use of these principles for segmenting scene text objects in video sequences with a relative motion between the camera and the scene. <br><br> A computer vision based system that can aid the pilot to detect obstacles in the flight path of an aircraft can be useful for avoiding collisions. Such a system would also be useful for development of a Synthetic Vision System (SVS) proposed for use in a High Speed Civil Transport (HSCT) aircraft with limited cockpit visibility. For this purpose, we had implemented a number of algorithms to detect airborne obstacles using image sequences obtained from a camera mounted on an aircraft. The performance of these algorithms was characterized in presence of camera noise using theoretical and experimental methods. Since the performance degrades in the presence of background clutter, a special approach to address the problem of hazard detection in presence of clutter was studied. This approach uses the differences in the behavior of translation and expansion of image features corresponding to the objects on a collision course and the background clutter. Algorithm fusion for combining different algorithms to overcome their individual limitations was also studied. In addition to this work on detecting objects on collision course, algorithms for detecting objects crossing the aircraft were designed and implemented on a real-time system. <br><br> Previous to our work on airborne object detection, we had worked on object detection on runways in presence of extraneous features, such as tire-marks. The work done on this topic is briefly described, citing references for details. The concepts used in this work were applied to another application: extraction of scene text in video image sequences with a relative motion between the camera and the scene. Assuming that the scene text generally occur on planar surfaces, the planar motion model is used to segment the scene into planar surfaces and determine their parameters. The clutter features which do not satisfy the planar motion model are removed as outliers. The parameters of each segmented surface are used to correct the perspective distortion of the surface. Each surface can then be analyzed separately for detecting and recognizing text objects, with perspective correction potentially improving the recognition performance.