Giving meaning to objects on video footage in the spatial-temporal unified framework

Processing video footage is a popular but challenging task today. With image processing methods, meaningful information can be extracted from moving and stable camera images. One of the most challenging aspects of working with images from moving cameras is the location of objects in an ever-changing environment. Besides, dealing with noise, reflection, scaling and similar challenges on video images is critical to developing a durable model. In the thesis study, studies were carried out on learning the motion directions of vehicles and also their motion angles using in-vehicle video images. In these studies, motion profiles obtained from video images were used. In this way, the image size was reduced, allowing the architecture to work in real time. YOLOv3, one of the successful deep learning architectures, was used to learn the motion directions of vehicles. In addition, a new architecture was created by developing YOLOv3 in the phase of learning the movement angles of vehicles as a new parameter. In this way, without the need for tracking algorithms, vehicles, motion directions, and angles can be learned on a single picture. The results obtained are at a level that can cope with current deep-learning architectures. In the presented method, bad weather conditions, noise in the image etc. factors do not adversely affect the result. Thus, a more consistent and durable system was obtained. Learning the movement angles and directions of the vehicles is important in terms of making sense of the movements of the vehicles and thus using them in anti-collision systems.