In-Car State Classification with RGB Images.

In the next years, shared autonomous vehicles are going to be a new reality. The absence of the human driver is going to create a new paradigm for in-car safety. This paper addresses this problematic by presenting a monitoring system capable of classifying the state of the vehicle interior, i.e. good or bad condition. We propose the use of classifiers, with RGB images, to infer the in-car cleanliness state. Moreover, 18 state-of-the-art classifiers were trained and evaluated, using pre-trained models. To be able to train and evaluate these approaches an in-car dataset was created with 3488 samples from 135 cars, and then split in 2439 train, 351 validation and 689 test RGB images. From all the evaluated, ResNet-18 showed the best results, achieving an average accuracy of 91.24% 123 Hz.