An optimized AdaBoost Multi-class support vector machine for driver behavior monitoring in the advanced driver assistance systems

摘要

Advanced Driver Assistance System (ADAS) is a Cyber-Physical System (CPS) application mainly developed for human–machine interaction. We employ the CPS information on ADAS while driving to not only improve the car’s driving conditionbut also alter the driving concepts to improve the functionality and safety of the vehicle. The driver can review the vehicle's feedback information in order to improve the ADAS system's ability to help in driving and completing the process of human–computer interaction. The data obtained during this interaction process are mainly high dimensional and consist of noisy/irrelevant features that pose an important obstacle to our ability to better grasp the car's driving states. In this study, we proposed an AdaBoost Multi-class Support Vector Machine (MSVM) with Cat Mouse Optimizer (CMO) algorithm for ADAS intrusion detection. The evolutionary algorithm is used in this research to tackle the drawbacks related with MSVM such as kernel, poor performance, minimum accuracy, overfitting issues, stability, and so on. Metaheuristics algorithms are an excellent and trustworthy solution for modeling MSVM parameters and producing optimal outcomes. As a result, we used the CMO method to optimize the MSVM parameters.From this, both normal and abnormal behaviors are predicted that affect the driving system.The experimental data is acquired from the power supply voltage data generated by the Jiangxi bus company. Data pre-processing is the major step to clean up, normalize and reshape the data for further processing. The low dimensionality features are extracted and finally, AdaBoostMSVM with CMO algorithm classifies both normal and anomalyactivities of driving vehicles. The results show that, in the majority of cases, the proposed model provides more accurate predictions in terms of accuracy, recall, precision, F-score, post hoc Nemenyi test, and Friedman test.Using performance metrics, the proposed method achieved 91.45 % accuracy, 94.45 % F-score, 93.89 % precision, and 94.90 % recall, outperforming other existing methods.