Enhancing Rule-Based Hands-Off Detection With Deep Learning and Permutation Feature Analysis
Soontae Kim, Kunsoo Huh · IEEE Transactions on Intelligent Vehicles · 2024
This paper explores the consistent and reliable hands-off detection performance to overcome the limitations of rule-based methods in determining driver's steering intent, a critical component of Level 2 autonomous driving systems. Driver steering states have been distinguished based on torque sensor analysis with steering dynamics models. In this study, limitations and key causes arising from differences in steering dynamics models are identified when applied to various vehicles, leading to accuracy degradation in commercialized vehicles. To address the irregularities between hands-on and hands-off characteristics across different vehicles, an LSTM (Long Short-Term Memory) based hands-off detection model is implemented first with additional inputs from forward cameras and in-vehicle sensors, some of which have not been utilized in existing rule-based methods. The Permutation Feature Importance technique is used to extract critical input features for hands-off detection, and the performance of the optimized model is verified to validate the importance of the detected features. Next, the newly identified significant signals are then used to enhance the existing rule-based method. Through experiments, the enhanced rule-based method is compared with the existing rule-based approach and, additionally, compared with learning-based and optimized learning-based methods. Finally, an optimized strategy considering resource constraints for implementation is discussed.