Detecting the vehicle presence by an embedded computer

Detekcija prisutnosti vozila na snimkama s kamere postavljene na urbanim raskrižjima u svrhu automatizacije upravljanja prometom u raskrižju predstavlja zanimljivu alternativu ugrađivanju induktivnih petlji. U sklopu rada ostvaren je i prikazan sustav za detekciju vozila u slikama raskrižja koristeći model pozadine. U radu je dan osvrt na srodnu literaturu. Predstavljeni su jednostupanjski i dvostupanjski modeli pozadine te metode izgradnje modela pozadine temljene na histogramu orijentacije gradijenta, mješavini Gaussovih razdioba i histogramu nijanse. Posebna pažnja posvećena je pristupima za ostvarivanje prilagodljivosti modela, ostvarivanju tolerancije na različite uvjete osvjetljenja te povećanju brzine obrade na ugrađenom računalu. Izvršen je niz validacijskih eskperimenata u svrhu optimiziranja hiperparametara opisanih metoda izgradnje modela pozadine. Ostvareni učinak različitih varijanti modeliranja pozadine te postupka detekcije vrednovan je na ispitnom skupu. Rezultati su predstavljeni nizom tablica i grafova te opširno komentirani.Detecting the vehicle presence in videos obtained by camera placed to record the urban crossroad with the goal of statistical data extraction and automatic traffic control makes an interesting alternative to existing systems of detection that use induction loops. This work explores the vehicle detection system implementation based on the background model. Related work concerning background modelling is discussed. Single stage and two stage background models are described along with background modelling methods based on histograms of oriented gradients, mixture of Gaussians and hue histogram. Special attention was given to approaches for realising adaptability of the model, achieving tolerance to different lighting conditions and increasing the detection process speed on the embedded computer. A set of validation experiments was conducted in order to optimise hyperparameters of described background modelling methods. The actual performance of detection system based on different background models was evaluated on the test set. Results are presented in a series of tables and graphs and discussed in detail