In this paper we demonstrate an approach to using a neural network for identifying observable 3D moving objects by their kinematic characteristics. Only coordinate measurements are used to identify an object. We suggest ResNet-like architecture (8 blocks and 283 469 learning parameters). Our solve good enough defines 13 classes of trajectories (12 object classes and 1 class "not trajectories"). We achieved 0:84 % correct answer on the F1 score on the only observed object.