Research On 3D Spatial Trajectory Clustering Algorithm Based On Spectral Clustering

In the traffic video detection system,clustering analysis of the motion trajectory becomes a hot research topic.There are many clustering algorithms were used to solve the problem of trajectory clustering in the complex traffic scene.However,some are based on the relationship between the feature corners which are in the 2D image plane,or some are based on the relative motion constraint relationship between the motion trajectories.These algorithms are deeply influenced by the camera perspective and light problems,thus they are not suitable for the detection in the complex traffic scene.In this paper,we study the motion constraint relationship of the feature points trajectories on the rigid body and propose a novel 3D trajectory clustering method which is based on the Spectral Clustering.Firstly,the motion trajectory is obtained by the feature extraction ORB algorithm and the bidirectional reversibility constraints KLT algorithm.Secondly,the similarity of the two trajectories is calculated by the motion constraint relationship of the feature points trajectories on the rigid body in 3D space,then the similarity matrix of the trajectory set can be constructed.Thirdly,Spectral Clustering algorithm is utilized to obtain the initial clustering results.Finally,the 3D parameters of the feature points are estimated based on the correspondence between the projection speed and the height of the feature points,then the 3D vehicle models are employed to deal with the inter-class merge processing with the initial clustering results,and the final clustering results can be got eventually.Moreover,the clustering algorithm is applied to different traffic scenes,and it has good performance.Multiple tests on some highway sections,the detection of the final clustering result can reach 95%.Meanwhile,the clustering algorithm can satisfy the real-time requirements of the video processing.