CONVOLUTIONAL NEURAL NETWORKS: AN APPROACH FOR VISUAL OBSTRUCTION DETECTION IN AUTOMOTIVE REVERSING CAMERAS

Abstract

In recent years, the study of Artificial Intelligence in the automotive industry has led to the design of intelligent systems applied to road safety, highlighting the importance of improving road safety worldwide, and thus reducing the number of accidents annually. One of the main functions of these systems is, for example, pedestrian detection, which is performed by cameras and radar-type sensors, among others. However, environmental factors cause visibility problems and obstructions that make pedestrian detection difficult and lead to collisions. With the purpose of contributing to the solution of the exposed problem, two case studies using Convolutional Neural Networks are applied in this research. The first using a pre-trained model (Inception V3) and the second, a proposed model (RvlsNet) to detect dirt on the lens of a vehicle's reverse camera. These types of factors directly affect visibility, which leads to an increased risk of collision when reversing the vehicle. Applying a general data mining methodology, we obtained a result of 0.9549 and 0.9416 accuracy, respectively, for the models used. Keywords: Convolutional Neural Networks, Classification, Obstruction, Detection, Reversing camera, Inception V3