20.500.12556/RUL-165972 Analiza pogojev in simulacija nastanka prometnih nesreč z metodami strojnega učenja Condition analysis and occurrence simulation of traffic accidents using machine learning Prometnih nesreč, še posebej takšnih s težjim potekom, najverjetneje ni mogoče povsem preprečiti - lahko pa jih poskušamo zmanjšati. Kot podporo odločevalcem, ki sodelujejo pri prometni varnosti, si številne ustanove prizadevajo najti najnevarnejše odseke cest in oceniti stopnjo tveganja. Cilj magistrske naloge je razvoj in evalvacija napovednih modelov za klasifikacijo nevarnih cestnih odsekov in ugotavljanje faktorjev, ki vplivajo na nastanek prometnih nesreč. V nalogi smo se osredotočili na razvoj celotnega cevovoda, ki je obsegal faze od prevzema in obdelave (prostorskih) podatkov, do implementacije in evalvacije različnih modelov. Velik poudarek smo namenili analizi prostorskih podatkov, s katero smo pridobili številne dodatne informacije o nesrečah in njihovi okolici. Ugotovili smo, da je za gradnjo napovednih modelov kritična faza ustvarjanje negativnih primerov. Pri eksperimentalnem delu se je najbolje obnesel model nevronske mreže, ki je dosegla 90% točnost in 81% priklic. Delo nakazuje, da lahko iz obstoječih podatkov s pomočjo napovednih modelov klasificiramo nevarne odseke in ugotovimo faktorje, ki najbolj vplivajo na nastanek prometnih nesreč. Kljub obetavnim rezultatom je za resno uporabo potrebno dodatno eksperimentalno delo, predvsem na področju ustvarjanja učnih primerov. Traffic accidents, especially severe ones, may never be completely avoidable, but their occurrence can be minimized. To support decision-makers in traffic safety, many institutions aim to identify the most dangerous road sections and assess risk levels. This master’s thesis focuses on developing and evaluating a predictive model to classify hazardous road sections and identify factors associated with traffic accident occurrence. The thesis outlines the creation of a comprehensive pipeline, covering stages from data acquisition and processing (specifically spatial data) to the implementation and evaluation of different models. A major focus was given to spatial data analyses, which provided valuable insights into the locations of traffic accidents and their surrounding conditions. In building predictive models, we found that creating negative samples of traffic accidents was essential for classification, which we generated in various ways. During testing, the neural network model performed best, achieving 90% accuracy and 81% recall. Our findings suggest that existing data can be used effectively with predictive models to classify dangerous road sections and determine the factors that most impact traffic accident occurrence. While the results are promising, additional experimental work is needed before practical application, particularly in refining training samples. napovedovanje klasifikacija prometne nesreče promet prostorski podatki modeli strojnega učenja forecasting classification traffic accidents spatial data machine learning models true false false Slovenski jezik Angleški jezik Magistrsko delo/naloga 2024-12-16 12:40:01 2024-12-16 12:40:08 2024-12-20 06:56:08 0000-00-00 00:00:00 2024 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 37099 219966211 Gercer_Matija_-_Analiza_pogojev_in_simulacija_nastanka_prometnih_nesrec_z_metodami_strojnega_uce.pdf Gercer_Matija_-_Analiza_pogojev_in_simulacija_nastanka_prometnih_nesrec_z_metodami_strojnega_uce.pdf 1 5D0C3CF9996C5DDB0004B06F6B1ADDD0 ee7cd626fa72d405475cf199606cb1191b8ce9a572767db890b2e412bca98483 fa06337a-bba1-11ef-b232-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=197307 Fakulteta za računalništvo in informatiko 0 0 0