Real-time motion trajectory training and prediction using reservoir computing for intelligent sensing equipment

Real-time moving target trajectory prediction is highly valuable in applications such as automatic driving, target tracking, and motion prediction. This paper examines the projection of three-dimensional random motion of an object in space onto a sensing plane as an illustrative example. Historical running trajectory data are used to train a reserve network. The trained network model is subsequently used to predict future trajectories. In the experiment, a network model trained on 20 000 frames of random running trajectory data was used to predict trajectories for 1–20 future frames, and 5000 frames were used for testing. The results showed prediction errors for 80% of the predictions of less than 0.01%, 0.8%, and 4% for 1, 10, and 20 future frames, respectively.