Results comparison

Each segment contains 10s (0-9) trajectory datapoints. We compared using different length of input sequence:

• 3s (2-4)
• 4s (1-4)
• 5s (0-4)

to predict 5s (5-9) trajectory into future.

Observation: Larger variance among predictions with lane change

5s input sequence gives the best performance. Hence, we compared adding extra features (vel_x and vel_y) into 5s input sequence.

Observations:

• Longer input sequence and more input features yield better performance
• Error values increase when the prediction horizon increases

The CARLA simulated data can be found here.