Application of binary classiï¬er by CNN in the detection of trafï¬c accidents

Abstract

The possibility that there are vehicles that travel without the need for a driver, that is, autonomous vehicles, is a glimpse of science fiction that in recent years has been gaining notoriety by major manufacturers and researchers. There is an imperative need to develop systems that can provide autonomous vehicles with the capacity for early collision identification with precision and safety, being an important topic hard explored by several areas, among which computer vision and artificial intelligence stand out. vast potential that both have presented when combined. Based on this need, this work aimed to employ computer vision and artificial intelligence techniques for the processing and classification of images extracted from short clips contained in recorded videos, thus obtaining a binary classification of a given situation that identifies the occurrence or not of an accident . Two architectures of convolutional neural networks were evaluated: AlexNet and ResNet-50, for the labeling of moments in a set of 19 videos, totaling 201 clips and 18,064 images analyzed in 30 epochs in the training phase. The effectiveness of the models was evaluated considering the measures F_1 score and Precision. The results were obtained in two different conditions: without improvements applied to the images and with improvements such as the histogram equalization. The results were: AlexNet F_1 score average of 91.5% against 89.5% of ResNet-50 for the first case and AlexNet F_1 score average of 88% against 91.5% of ResNet-50 for the second.