Online Driving Style Analysis Based on Genetic Algorithm and CatBoost

Accurate driving behavior analysis can not only greatly improve driving safety and driver comfort, but also facilitate the development of trajectory planning, abnormal driving detection and other applications. The previous driving style analysis required the complete timing data of the driver, which could not be extracted in real time. This paper presents an online driving style analysis method, which is mainly divided into two parts. The first part is driving data segmentation and clustering based on K-means and genetic algorithm. Different subsequences are obtained by sliding window segmentation. The total KL divergence within the cluster is taken as the objective function, and the sub-sequence clustering results are obtained by genetic algorithm. Then, according to the final clustering results, the driving style of drivers is measured from three dimensions: statistical index, cospin similarity and KL divergence. The second part is the online driving primitive extraction model. According to the clustering results (labeled data) obtained in the first part, different machine learning models are trained, and the performance of random forest, KNN, SVM and CatBoost models is compared. Through the trained machine learning model, we are able to extract driving primitives in real time. The experimental results show that the first part can extract driving primitive and analyze driving style, and the second part CatBoost achieves better results than other machine learning models, with prediction accuracy of 0. 8, recall rate of 0. 82 and f1_score of 0. 81.