Is it possible to do motion prediction with perception dataset

[ahsan155]

I am using the waymo perception dataset v1.4.2 that comes with map. The tfrecord file has a field: laser_labels { box { center_x: 39.862915373549185 center_y: 37.57924021856002 center_z: 3.9028802750369778 width: 2.07522034497737 length: 3.6631732945601687 height: 1.7400000000000038 heading: 1.563768235771243 } metadata { speed_x: -8.445226685916751e-29 speed_y: 0.0 accel_x: 1.0556533986614397e-28 accel_y: 0.0 speed_z: 1.6494583370931155e-31 accel_z: -2.0618230442606244e-31 } type: TYPE_VEHICLE id: "-HJlbwWGxigWVjAFBBdTxA" num_lidar_points_in_box: 52 most_visible_camera_name: "FRONT_LEFT" camera_synced_box { center_x: 39.859376980729586 center_y: 37.582941913788076 center_z: 3.902378373465603 width: 2.07522034497737 length: 3.6631732945601687 height: 1.7400000000000038 heading: 1.563768235771243 } num_top_lidar_points_in_box: 53 }

I want to do motion prediction with vehicle data. To do motion prediction I need vehicle position, velocity and heading in the correct frame. I found out that above information is in local frame. I already transformed box center_x and center_y in global frame. If I want to do motion prediction, do I also need to transform metadata speed_x/speed_y and box heading to global frame. If so, how would I do it.

ground truth trajectory (orange) and predicted trajectory (red)

Predicting trajectory in this scenario should be simple as trajectories are mostly horizontal. What could be a possible explanation that predicted trajectories are making turn like that. Is it because of not transforming heading/velocity?

I trained motion prediction model with motion dataset. I want to get all the necessary information from perception data, store them in a tfrecord file and run it through prediction model. Overall, I wanted to know if motion prediction is possible with all the correct transformation.

[scott-ettinger]

Something may be going wrong with your transformation of the trajectories. The example code here shows how to do this transformation. In general, you should be able to align the trajectories with the map data.