D.J. Zhang, F.M. Shang

Pages: 15-28

Abstract
Accurately and quickly identifying harmful driving behaviors is of great research significance. This article introduces a multi class Adaboost algorithm for accurately and quickly identifying harmful driving behaviors on urban roads. This method first carefully collects data related to bad driving behavior on road sections. Then, the optical flow method is used to capture images of driving behavior, and the TV-L1 optical flow method is used to refine the images to improve clarity. Finally, the CART decision tree serves as the cornerstone for constructing a rigorously trained weak classifier, strategically combining multiple individual weak classifiers to form a robust strong classifier. This powerful classifier is supported by the Adaboost algorithm and can provide detailed classification of bad driving behavior. To ensure a holistic assessment, fuzzy mathematics is deployed to discern behavioral risks, thereby facilitating the correlated identification of adverse driving patterns. Experiment results verifies the efficacy of our approach, with recognition time dwindling to a mere 1.9 seconds and the accuracy of correlated recognition soaring to an impressive 99.8%.
Keywords: multi class Adaboost algorithm; bad driving behavior; association recognition; urban RD; cart decision tree