Applying continuous profile models for analysis of multidimensional trajectory clustering results

Problem of clustering trajectories is pre-conditioned by a need to organize motion of objects under control. The polynomial regression method is the best approach to trajectory cluster selection, which estimates a form of general trajectory of each cluster. For a set of sufficiently heterogeneous trajectories, the defined clusters are heterogeneous also. For the inhomogeneous cluster, its polynomial regression is a too strong abstraction. To demonstrate the cluster heterogeneity (and, thus, non-full clustering trajectories) a method of dimension reducing is in need. So, linear dynamical system models are applied to object multi dimensional trajectory clustering by the polynomial regression method. Extracted components of system state vector give an opportunity to elicit cluster fine structure in accordance with cosine measure. In present paper applying continuous latent profile models is shown for determining a central trajectory in sufficiently homogeneous clusters of multi-dimensional trajectory projections. Continuous profile models are the hidden Markov models with internal states of time and scale. Defining the cluster central trajectory with affinity estimation by cosine measure in use gives an opportunity to determine the cluster outlying trajectories in potential, which represent motions located in a risk zone.