Demiryolu sürücü destek sistemleri için yaya saptama

Pedestrian detection is one of the most studied issues of advanced driver assistance systems. Although a tremendous effort is already given to create adequate datasets and to develop advanced classifiers for cars, studies about railway systems remain very limited. The research done within the scope of the thesis shows that direct application of neither existing advanced object detection systems, nor specifically created ones for pedestrian detection (such as classifiers which is pre-trained well-known pedestrian datasets Caltech, INRIA etc.), can provide enough performance to overcome railway specific challenges. Fortunately, it is also shown that without waiting the collection of a mature dataset for railways as comprehensively diverse and annotated as the existing ones for cars, a transfer learning approach to fine-tune various successful deep models (pre-trained using both extensive image and pedestrian datasets) to railway pedestrian detection tasks provides an effective solution. In the light of this information, to achieve transfer learning, a new Railway Pedestrian Dataset (RAWPED) is collected, annotated and divided into challenge based subgroups. Moreover, the localization and adaptation limitations of deep models are resolved with a feature-classifier ensemble. The application of resulting two stage system to various railway scenes demonstrate that employed transfer learning strategies enable reliable adaptation of pre-trained models to railway pedestrian detection scenes. Furthermore, complementary properties of the transferred models, classifiers and diversity of their results are analyzed. Based on the findings, a novel machine learning strategy is structured to create an ensemble, which defragments outputs of individual models and performs consistently better than its components.