Abstract: The vehicle trajectory extracted from highway surveillance system can be used to analyze and recognize vehicle behavior.Due to a small amount of abnormal trajectory,such as change lanes and overtaking,the classic spectral clustering with longest common sub-sequence(LCSS) can’t effectively distinguish all kinds of trajectory.In addition,the popular HMM trajectory model ignores the negative impact of the samples and only classifies them by maximum likelihood value to cause a higher rate of false recognition in vehicle behavior recognition.In order to address these issues,according the characteristics of highway vehicle trajectory,we proposed a vehicle trajectory recognition method based on quadratic spectral clustering and HMM-RF hybrid model.Firstly,the trajectory curvature is calculated to distinguish overtaking by curved characteristics,and then lane changes trajectory is distinguished by spectral clustering with inclination similarity in non-curve trajectory.Secondly,all the sub-clusters are clustered by spectral clustering with LCSS again,which can effectively distinguish overtaking,changing lanes and normal trajectory.We made the output of HMM model,the different dimension of probabilities as an input for random forest model to improve the precision of behavior recognition.We did experiments under different data sets to verify the effectiveness of the method.The average accuracy rate of trajectory clustering can achieve 96%,and the average accuracy rate of behavior recognition can reach to 89.3%,so the algorithm has higher accuracy and robustness.

Key words: Trajectory clustering,Vehicle behavior recognition,Quadratic spectral clustering,HMM-RF hybrid model