Accelerated gradient methods for nonconvex optimization: Escape trajectories from strict saddle points and convergence to local minima