Comparación de redes Vision Transformer y convolucionales para detección de conducción segura

This paper presents the results of comparing the training of deep learning architectures applied to the development of safe driving systems. Databases were generated with 670 images of drivers inside vehicles, which were divided into three subsets for training two architectures based on convolutional neural networks (CNNs) and transformer networks for vision. 70% of the images were used for training, 20% for validation, and the remaining 10% for testing. These two architectures were compared to assess their pattern recognition capabilities in classifying three driving states, normal state, distracted state and sleep state. In both cases, the need to focus the learning to improve the learning performance of the two architectures is evident, for which a previous stage of face segmentation by means of Haar classifier is included, obtaining accuracy levels of 98% for the CNN and 87% for the Transformers network with average inference times of 0.1 and 0.52 seconds, F1 scores of 98.9% and 82.2%, and recall rates of 98.8% and 80.6%, respectively, the statistical metrics for each class demonstrate a high degree of confidence in the recognition of each class. The comparison was performed on a computer with a 2.3GHz Core i9 processor, 24GB of RAM, and an RTX 4080 GPU with 12GB of memory, using MATLAB programming software.