This paper presents a novel supervised features-based learning method to classify particle dynamics in biological imaging. To this goal, we consider geometric features computed on trajectories and encoding their intrinsic geometrical characteristics. The method is validated on a dataset simulating different processes: Brownian motion, directed Brownian motion, Orsntein-Ulhenbeck process, fractional Brownian motion, and Continuous-Time Random Walk. The presented approach allows for distinguishing these five dynamical behaviors in a unified framework with high accuracy, and its strength lies in distinguishing several subdiffusive dynamics from free or superdiffusive ones. The robustness to image noise and trajectory length variation is proven, showing the flexibility and reliability of the method in terms of variability due to acquisition techniques. Distinguishing different subdiffusive behaviors strongly impacts particle analysis in biology, and an application is shown to the motion classification for receptors (CCR5) at the cell membrane.