Detección de incidentes automovilísticos usando técnicas de aprendizaje de máquina

Abstract

The increase of automobiles on the roads has caused several types of problems, among the most serious are traffic incidents, frequently caused by speeding. These directly influence the risk of an impact collision, as well as the probability of death and serious injury. Every year around 1.35 million deaths are registered worldwide due to traffic accidents and around 50 million get injuries as a result of those accidents; originating an annual cost of 3% of GDP to the government of each country. In recent years, Artificial Intelligence has become a fundamental part of autonomous systems in the field of transport because it helps control them with minimal human intervention, as in the case of video surveillance on the roads. It is important to develop automation tools that help security camera operators to detect traffic incidents, because for every minute saved in early detection of traffic incidents, up to five minutes can be gained in the road recovery phase. The present work has created and compared 100 classification models such as kNN, logistic regression and SVM. A dataset was created from videos from UCF and CADP which was used to calculate the estimated speeds using detection and tracking algorithms. These speeds were used for the training and evaluation of the models, obtaining a model with an accuracy of 0.95 and AUC ROC of 0.96.