Distributed edge computing system for deep learning based real-time video analysis applications

Deep learning studies and, accordingly, artificial intelligence systems have become an area of great interest in our country. Especially the studies in the last 5 years have brought to light the deficiencies in this field. There is a need for high cost equipment for studies in this field. There is also a lack of dataset specific to Turkey. With this thesis, datasets specific to Turkey related to license plates, traffic signs and traffic lights have been prepared and their applicability has been demonstrated by making improvements on appropriate deep learning models. In this context, pre-trained deep learning model APIs on detection and recognition of traffic signs and traffic lights were used. TensorFlow, Keras deep learning libraries are used for these models. Pre-trained Faster R-CNN and SSD models with TensorFlow Object Detection API using these libraries were trained with the datasets we prepared. A total of 6 datasets specific to Turkey were prepared, respectively, 2 for vehicle license plate recognition, 1 for vehicle license plate detection, 1 for traffic sign detection and recognition, and 2 for traffic lights detection and recognition. It is an important problem that the datasets we have prepared in the industry sector and other sectors and the models trained with them can be run in real time with cost-effective hardware. For this, it has been tried to show the applicability of the models with edge computing methods on edge devices by using cost-effective edge device and video processing unit (VPU). It has been ensured that different models are run on edge devices with distributed edge computing methods using video images. With this thesis, it has been shown how to prepare data sets in accordance with the format, size and color standards in Turkey in an original way. It has been shown that suitable deep learning models trained with these prepared datasets can be developed as a distributed artificial intelligence system on edge devices in real time.