Enhanced Nvidia-based deep learning model with inverted residual blocks for self-driving car

This paper presents a novel architecture with an aim to improve steering angle prediction in autonomous driving systems by using the inverted residual blocks which are inspired from the MobileVNet in the original Nvidia architecture. The proposed model processes input images sized at 66 x 200 x 3 (height, width, and RGB channels), effectively enhancing the feature extraction and also minimizing the risk of vanishing gradients. The dataset that was used in this study is extracted from the Udacity self-driving car simulator which consists of over 24,591 images coupled with their corresponding steering angles. The proposed model was trained on data that was heavily pre-processed and various data augmentation techniques were used to ensure strong performance in various driving scenarios. The proposed model achieved an impressive MSE of 0.03210 and an R² value of 0.9410 whereas the traditional NVIDIA model achieved an MSE of 0.05710 and an R² value of 0.8960, and MobileNetV2 model showed an MSE of 0.0988 and an R² value of 0.7941 which showcases the proposed model's impressive capability of accurately predicting the steering angles. This unique approach reflects the importance of inverted residual blocks in behaviour cloning models to improve the model generalization and convergence ability. The findings of this study present a valuable contribution to the field of deep learning applications in the automotive industry, encouraging further exploration of lightweight, efficient model architectures in the quest for smarter transportation solutions.