search

autowarefoundation / autoware.universe

https://autowarefoundation.github.io/autoware.universe/
Apache License 2.0
967 stars 631 forks source link

Misdetects the pedestrian near a car #7563

Open sgNicola opened 3 months ago

sgNicola commented 3 months ago

Checklist

Description

The perception system fails to detect and track the pedestrian when they are close to the car. It only recognizes the pedestrian once they move away from the vehicle. image walker Here is the video of the rviz2. https://github.com/autowarefoundation/autoware.universe/assets/115910640/b538ca29-b0c7-4f02-a50a-64bf7100e8c2

Expected behavior

Detect the pedestrian correctly

Actual behavior

Miss detects the pedestrian near the car.

Steps to reproduce

  1. cd autoware/src/sensor_kit/external/awsim_sensor_kit_launch/awsim_sensor_kit_launch/launch，modify the awsim_sensor_kit, only one top lidar is used. image
  2. ros2 launch autoware_launch logging_simulator.launch.xml map_path:=map_path vehicle_model:=sample_vehicle sensor_model:=awsim_sensor_kit
  3. replay this bag https://drive.google.com/file/d/1FFE1XTrY-A8ELE1zi32bvGfe-Hah4u7P/view?usp=sharing

Versions

-OS: Ubuntu 22.04 -ROS2: Humble -Autoware: Autoware.universe main

Possible causes

The point clouds of pedestrians and cars are connected together， and not recognize correctly

Additional context

No response

badai-nguyen commented 3 months ago

@sgNicola Thank you for your issue report. I think the current DNN model (Centerpoint) is not good enough for pedestrians nearby the vehicle, especially for the large vehicle like bus because of lack of training dataset. Could you try to enable detection_by_tracker for pedestrians by this and check this again?

@YoshiRi Do you have further comment?

amadeuszsz commented 2 months ago

Out of curiosity, I tried with TransFusion, hoping intensity field will play main role here. Despite the fact the pedestrian and car results with different intensities, it still brings the same result as CenterPoint. transfusion_car_with_pedestrian

For TransFusion It could be due to different intensity profile in input data. I see Velodyne here, AFAIK dataset consist of Velodyne data as well. Therefore, the source for both ML models is lack of similar samples in dataset as @badai-nguyen mentioned.

stale[bot] commented 3 weeks ago

This pull request has been automatically marked as stale because it has not had recent activity.