Predicting Human Trajectories in Multi-Class Settings

Future Work Path Error = Distance at each predicted step between prediction and ground truth. Final Displacement = Distance between the final points ● We used 5 pre-annotated Stanford Drone Dataset video-streams of the “The Circle of Death” ● total of 22 minutes of footage (with 1548 Bicyclists, 917 Pedestrians, 107 Skaters, and 132 carts) ○ Holdout: Training on first 3 scenes (17 min). Tested on 4th scene (5 min) ○ Training Set = 1307 Bikes, 831 Pedestrians, 91 Skaters, 111 Carts Naive GRU Model: Assume Independent trajectories (Basic Sequence Generation Problem) GRU: learning rate = 0.003 (with annealment), hidden state dim. = 128, 2 GRU layers, embedding with ReLU nonlinearity. Using Mean Squared Error & no BPTT truncation.