Dangerous Driving Behavior Recognition Based On Projection Twin Vector Machines

Dangerous driving behavior is currently one of the most important factors leading to traffic accidents.As two typical dangerous driving behaviors,fatigue driving and aggres-sive driving not only threaten people’s life and property safety,but also seriously hinder the development of transportation industry.Therefore,using machine learning methods to accu-rately identify driving fatigue and driving style has become a research highlight.This thesis studies the above dangerous driving behaviors from the perspectives of multi-view learning and semi-supervised learning,which effectively improving the accuracy of driving fatigue detection and driving style recognition.The main works are as follows:（1）This thesis deeply analyzes the current research status of driver fatigue detection and driving style recognition.Machine learning algorithms such as Support Vector Machine are summarized in terms of their principles,solving methods and classification criteria.The current multi-view and semi-supervised learning algorithms are deeply studied,which lays a theoretical foundation for the research of dangerous driving behavior recognition.（2）To solve the problem of driving fatigue detection,this thesis proposes two new multi-view classification models:Multi-view Projection Twin Vector Machine（Mv PTVM）and its robust variant（RMv PTVM）.The models can integrate complementary information from different views,and find multiple projection vectors for each class of samples to achieve well separation of different classes in the projected space.RMv PTVM replaces the square L₂norm metric in the objective function with a more stable L_2,1norm to improve the robust-ness to outliers.Experimental results on driving fatigue data and UCI datasets show that the proposed algorithm has better performance.（3）For the problem of driving style recognition,this thesis proposes a Semi-Supervised Projection Twin Vector Machine（SSPTVM）base on the semi-supervised learning frame-work.By taking full advantage of the characteristic information encoded from the original multivariate sensor data by Hyperdimensional Computing（HDC）,SSPTVM can optimize a pair of projection matrices to classify samples.In addition,the particle swarm optimization（PSO）algorithm has been improved to meet the demand that the optimized model param-eters contain discrete variables.Experimental results on real driving style data confirm the effectiveness of our method.In conclusion,the dangerous driving behavior recognition models proposed in this thesis effectively improves the accuracy of driving fatigue detection and driving style recognition,providing more reference schemes for improving the level of traffic safety and the develop-ment of intelligent driving in the future.