A Comparative Study on End-to-End Learning for Self-Driving Cars

Abstract

Autonomous vehicle technology has advanced in recent years. The self-driving car is one of the most attractive research fields, and automakers are fast focusing on it. There have been a number of attempts made in this field, such as lane recognition, the detection of objects on roadways, and the reconstruction of three-dimensional models; however, the focus of our study is on models that directly transform the camera input images into steering angles. In this paper, we performed a comparative study of some of the popular end-to-end CNN models pertaining to autonomous vehicles. We used four different data sets for model training and validation. Only one of the data sets was gathered from the real world; the other three were created using software simulations. For evaluating the performance of different models, we used the mean squared error (MSE) metric. It was interesting to see that certain models fared better than others when applied to diverse data sets. When considering real-world datasets, both pre-trained VGG-16 and pre-trained VGG-19 using transfer learning exhibit comparable performance, achieving an MSE value of 21.4 which is better than all other considered models. However, in the case of simulated datasets, pre-trained VGG-19 outperforms the majority of the other models. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.