A neural network-based lane-keep assist (LKA) function: A sensitivity analysis related to prediction accuracy and false positives

Abstract

Road accidents are one of the leading causes of deaths worldwide (Road traffic injuries 2022). Around 1.3 million fatalities were reported due to road traffic crashes, and the injuries and economic losses are even higher. The cause of the majority of these car accidents can be attributed to human errors while driving (research note: 2016 fatal motor vehicle crashes: overview - transportation 2017), emphasizing the importance of reliable safety systems in modern automobiles. Lane Keep Assist (LKA) is one such system commonly referred to as an Advanced Driver Assistance System (ADAS). The LKA system plays a critical role in ADAS-equipped vehicles by assisting vehicles in maintaining their lane positions. This research explores various factors that affect the performance of a neural network-based LKA system. Real-world driving data, including ego vehicle states and environmental information captured through a forward-looking camera, has been collected and been made available for this study. The collected data then underwent processing and normalization as part of this study, to facilitate subsequent machine learning and analysis. To enhance the dataset’s variability and improve the performance of the machine learning models, various data augmentation techniques were employed. The augmented data, along with an appropriate sample size, was then used to train different machine learning models. A main objective was to determine the optimal combination of data sampling, data augmentation, and machine learning algorithms. The evaluation of the models is based on multiple metrics, with a primary focus on intervention prediction accuracy. This metric measures the system’s ability to accurately predict the need for LKA intervention based on the input signals provided. Additionally, the Area Under Curve of the Receiver Operating Characteristics curve (AUC-ROC) is used as a secondary evaluation metric. Furthermore, tools such as the confusion matrix are utilized to obtain a visual representation of the system’s performance. The findings of this study provide valuable insights into the influence of a variety of modeling parameters and methodological choices on the performance of neural networkbased LKA systems