Enhancing Accuracy and Robustness of Steering Angle Prediction with Attention Mechanism

TL;DR

Mean squared error ($MSE$) is the evaluation metric for steering angle prediction, defined as $MSE = (1/n) \sum_{i=1}^n (r_i - r_i\_hat)^2$. The paper investigates compact ResNet and InceptionNet backbones and introduces an attention module to improve both accuracy and adversarial robustness in steering angle forecasting. Empirical results show attention yields up to 6.83% MSE reduction on a Kaggle SAP dataset and 6.09% on a custom Udacity-derived dataset, with robustness improvements under FGSM/PGD attacks reaching up to 53.95% and 56.09% respectively. These findings highlight the practical value of attention in real-world autonomous driving systems and include a publicly available dataset to facilitate further research.

Abstract

In this paper, our focus is on enhancing steering angle prediction for autonomous driving tasks. We initiate our exploration by investigating two veins of widely adopted deep neural architectures, namely ResNets and InceptionNets. Within both families, we systematically evaluate various model sizes to understand their impact on performance. Notably, our key contribution lies in the incorporation of an attention mechanism to augment steering angle prediction accuracy and robustness. By introducing attention, our models gain the ability to selectively focus on crucial regions within the input data, leading to improved predictive outcomes. Our findings showcase that our attention-enhanced models not only achieve state-of-the-art results in terms of steering angle Mean Squared Error (MSE) but also exhibit enhanced adversarial robustness, addressing critical concerns in real-world deployment. For example, in our experiments on the Kaggle SAP and our created publicly available datasets, attention can lead to over 6% error reduction in steering angle prediction and boost model robustness by up to 56.09%.

Figures (7)

• Figure 1: Illustration of deep-learning-based steering angle prediction on visual data. The deep model is designed to analyze visual data captured by the left, central, and right cameras, facilitating the inference of steering angles.
• Figure 2: Comparison of ResNet32 and ResNet32 with attention (two models are of similar sizes). The attention module consists of two basic branches: the trunk branch and the attention mask branch. More details in Sec. \ref{['sec:attentionmodule']}.
• Figure 3: ResNets vs. InceptionNets in terms of MSE loss vs number of parameters on the Kaggle SAP dataset.
• Figure 4: MSE comparison between ResNet32 vs ResNet32 with attention modules on the Kaggle SAP dataset.
• Figure 5: MSE comparison between ResNet26 vs ResNet26 with attention modules on the Custom dataset.