Person Detection in Images

Detekcija osoba u slikama ima široku primjenu u različitim područjima. U ovom radu prvo je dan pregled nekih tehnika koje su se za detekciju koristile kroz povijest, a najviše je pažnje posvećeno konvolucijskim neuronskim mrežama. Prvo je teoretski opisano kako one funkcioniraju, a potom je dan uvid u gradnju arhitekture i učenje pomoću biblioteke Keras. U nastavku su prokomentirani dobiveni rezultati i opisane vrijednosti na koje moramo pripaziti, kako ne bi došlo do pretreniranosti. Predikcija je izvedena pomoću tehnika "klizeći prozor" i "selective search" te su te dvije tehnike uspoređene.Person detection in images has wide application in many different fields. First thing mentioned in this thesis are techniques that were used for person detection through the history. Most of attention is given to the concept of convolutional neural networks. In the beginning it is described how convolutional neural networks work. Also, it is described how architecture for such networks is built and how the process of learning works. Afterwards, the results are presented and thoroughly commented as well as values which must be monitored to avoid overfitting. The prediction is performed with the "sliding window" and "selective search" techniques