Vehicle Detection and Tracking on RGB and Hyperspectral Data

Postgraduate Thesis uoadl:2073256 435 Read counter

Unit:
Κατεύθυνση / ειδίκευση Τεχνολογίες Πληροφορικής και Επικοινωνιών (ΤΠΕ)
Πληροφορική
Deposit date:
2017-10-26
Year:
2017
Author:
Filippas Dimitrios
Supervisors info:
Σέργιος Θεοδωρίδης, Καθηγητής, Τμήμα Πληροφορικής, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Original Title:
Εντοπισμός και Παρακολούθηση Κινούμενων Οχημάτων από RGB και Υπερφασματικά Δεδομένα Βίντεο
Languages:
Greek
Translated title:
Vehicle Detection and Tracking on RGB and Hyperspectral Data
Summary:
New trends on the scientific field of computer vision instruct for the application of advanced methodology techniques towards the exploitation of the rapidly grown available abundance of remote sensing data. To this end, a generic framework towards multiple moving vehicles tracking in RGB and hyperspectral video sequences was designed, developed and validated. Video data were collected on a specific spot of “Attiki Odos” motorway. Targeted objective of this work was the specific study of techniques for processing the hyperspectral video, as in this case, the problem of detection is more complicated and demanding compared to the RGB video. The initial pre-processing steps included frame registration and projective transformation for the RGB video, and also pixel interpolation and projective transformation for the hyperspectral video. In the first stage of the methodology, background subtraction is performed. The background subtraction for the RGB video was achieved employing an adaptive background subtraction technique. For the hyperspectral video a fusion of two techniques based on adaptive background subtraction and three-frame differencing, was employed. The second stage includes the detection of moving vehicles applying cost-minimization and, finally, the vehicle-tracking as long as they are recorded. It is worth noting that from the available bands in the hyperspectral video, several multi-band images were considered. The proposed methodology incorporates also certain rules that eliminate the effect of noise on the detection step and thus significantly improved its performance. These rules set restriction on the allowed range of values of certain parameters related to the size and position of vehicles, their direction and change of motion, and the time length and sequence of their recording in the video. The performed quantitative evaluation indicated that the developed approach was validated with overall accuracy rates of 84\% on a 1000-frame RGB video. For the evaluation of the methodology on hyperspectral data various experiments using different band combinations were conducted on a 500-frames video. The most successful combination, including three hyperspectral bands, was validated with overall accuracy rates of 79\%. Concluding, the goals set for this study are considered fulfilled since the most challenging objective of accurate moving vehicle tracking on hyperspectral data was achieved at a satisfactory degree.
Main subject category:
Science
Keywords:
Vehicle detection, Vehicle tracking, Hyperspectral video, RGB video, Background Subtraction, Kalman Filter
Index:
Yes
Number of index pages:
2
Contains images:
Yes
Number of references:
29
Number of pages:
104
document.pdf (2 MB) Open in new window