Khaled Adel Ezzat <br/><br/>
EVOLUTION OF OBJECT DETECTION, TRACKING, AND MOTION ESTIMATION ALONG WITH DEEP NEURAL NETWORKS/ 
Khaled Adel Ezzat 
 -  2022 
 - 61 p.   ill.   21 cm.  
<br/><br/>Supervisor: <br/>Khaled Foad<br/>Mustafa Elattar<br/>  <br/>Publication:<br/>1-On the Application of Hierarchical Adaptive Structured Mesh" HASM®" Codec for Ultra Large Video Format<br/>https://dl.acm.org/doi/abs/10.1145/3436829.3436870<br/><br/>2-nnDPI: A Novel Deep Packet Inspection Technique Using Word Embedding, Convolutional and Recurrent Neural Networks<br/>https://ieeexplore.ieee.org/abstract/document/9257912<br/><br/>3-On Optimizing the Visual Quality of HASM-Based Streaming—The Study the Sensitivity of Motion Estimation Techniques for Mesh-Based Codecs in Ultra High Definition Large Format Real-Time Video Coding<br/>https://link.springer.com/chapter/10.1007/978-981-33-6129-4_15<br/><br/>4-MinkowRadon: Multi-Object Tracking Using Radon Transformation and Minkowski Distance<br/>https://ieeexplore.ieee.org/abstract/document/9581542 
<br/><br/>Thesis (M.A.)—Nile University, Egypt, 2022 . 
<br/><br/>"Includes bibliographical references" 
<br/><br/>Contents:<br/>Chapters:<br/>1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1<br/>1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2<br/>1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5<br/>1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9<br/>2. MinkowRadon Multi-Object Tracking Technique . . . . . . . . . . . . . . 12<br/>2.1 Dataset - MOTChallenge (MOT20 & MOT17) . . . . . . . . . . . 13<br/>2.2 Model Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 14<br/>2.3 Numerical Results and Discussion . . . . . . . . . . . . . . . . . . . 20<br/>2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23<br/>3. The Adaptive MinkowYolo Multi-Object Detector-Tracker . . . . . . . . 24<br/>3.1 Dataset - MOT + Visual Tracker Benchmark (VTB) . . . . . . . . 26<br/>3.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 27<br/>3.3 Numerical Results and Discussion . . . . . . . . . . . . . . . . . . . 37<br/>3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42<br/>4. Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . 43<br/>Bibliography . . . 
<br/><br/> Abstract:<br/>Object Detection, Tracking and Motion Estimation have been a major concern<br/>since the 1970s, from Self Driving Cars, Surveillance Cameras, Industrial robotics,<br/>Traffic monitoring, Medical diagnosis systems, to Activity recognition, are expecting<br/>a huge increase in demand for automated detection-tracking systems. Modern hardware<br/>specifications and evolving deep learning applications with advancement of Computer<br/>Vision and Digital Video Processing are resulting in a massive progress towards<br/>fully automated systems, with all advance models and systems like R-CNN, YOLO,<br/>SSD, and RetinaNet, there will always be a trade-off between precision (mAP) and<br/>speed (FPS) which puts a new limits to computer vision advancement. Technological<br/>merging has the potential to drive the intuition to achieve such advancements, and<br/>overcome some of the existing limitations. Introducement of a combination between<br/>Deep Neural Networks and Digital signal processing to enable once again progress<br/>to be done in improving Object Detection, Tracking and Motion Estimation in a<br/>real-time videos. Utilizing both of the fields, this thesis purposes a complete detection/<br/>tracking framework utilizing YOLO v4 as a state-of-art object detector to<br/>detect the objects in the video sequences. In addition to a novel MinkowRadon tracking<br/>algorithm which utilizes the Radon Transformation and Minkowski Distance to<br/>translates the rest of video frames sequence to the signal’s domain, in an attempt<br/>iv<br/>to tackle extreme object tracking problems found in video sequences like eg. trembling<br/>moving cameras, deformation, motion blur, fast motion, and in-plane rotation.<br/>Tracking through signals have proven with a higher accuracy compared to the stateof-<br/>art tracking techniques that a combination between classical techniques and deep<br/>learning models is sufficient to solve most modern problems. 
<br/><br/><br/>Text in English, abstracts in English . 
<br/><br/><label>Subjects--Topical Terms: </label><br/>Informatics-IFM
<br/><br/><label>Index Terms--Genre/Form: </label><br/>Dissertation, Academic
<br/><br/><label>Dewey Class. No.: </label>610