Trajectory learning and analysis based on kernel density estimation

This paper presents a novel kernel density estimation approach to vehicle trajectory learning and motion analysis. The framework comprises a training stage and a testing stage. In the training stage, vehicle trajectories are first clustered by the hierarchical spectral clustering method. Then, through the proposed kernel density estimation approach, the average kernel density of one point on a trajectory can be estimated. In the testing stage, the compactness estimated by a Gaussian kernel function is introduced. Abnormal trajectories are detected with compactness lower than expected for a few consecutive frames. Vehicle motions are identified into multiple activities with their respective trajectory compactness.