HAAR-benzeri öznitelik temelli rastgele karar ormanı ile tek seferde obje saptama ve sınıflandırma

Object detection and tracking have been studied for decades and many algorithms have been introduced. Vision-based object detection and tracking became an important task with the increasing number of surveillance cameras. However, false alarm rates are still an issue to be solved in human operator managed scenarios. As precision and accuracy increase, false alarm rates become more manageable. In this thesis, a novel system for single-shot detection and classification of the object in images is introduced. For this purpose, we implemented Random Decision Forests (RDF) using Haar-like features. RDF and Haar-like feature calculation implemented on GPU are known for their test time speed. Thus, we are using RDFs for pixel level object classification, a methodology known for its balanced test-time performance both for speed and quality. The increase in accuracy is shown by conducting experiments on MNIST, INRIA, and PETS09 data-sets. As a demonstrative application, we used proposed RDF for on-road vehicles detection and tracking. A Sequential Monte Carlo method based algorithm, also known as Particle Filter (PF), is implemented for tracking detected objects. For non-linear and non-Gaussian processes, PF is a powerful methodology and is easy and preferable to be implemented on GPU with RDF. The proposed system puts emphasis on real-time speed of the algorithm on conventional computers. Compared to YOLO (You Only Look Once), our method shows comparable vehicle detection accuracy and computational speed in a conventional computer. Moreover, we are introducing a new framework where different tracking algorithms can be implemented and tested. It provides various modules for data extraction, data generation, training and testing algorithms with different parameters. Usage of the modules in the framework is also discussed.