Replace Traffic light detection model

[badai-nguyen]

Checklist

• I've read the contribution guidelines.
• I've searched other issues and no duplicate issues were found.
• I've agreed with the maintainers that I can plan this task.

Description

• Current Trafic Lights (TLs) fine detection model operates using cropped image obtatined via rough ROIs. It requires following input topics:

  • image_raw
  • camera_info
  • rough/rois
  • expect/rois

• traffic_light_recognition pipeline
  

• Limitations of the current approach

  • The fine detection is executed only after traffic_light_map_based_prediction which causes unnecessry delays.
  • The fine detector relies only local image of TLs, lossing global feature such as vehicles, road related features which might increases false-positive result when some vehicle's lamps included in rough rois regions.
  • This significantly relies on autoware_traffic_light_occlusion_predictor which is high computation cost and cause heavy bandwith of raw pointcloud transfer.
  • It might be a challenge to intergrate with other open-source TL detection models such as YOLOX.

Purpose

The task is open to update TLR pipeline to address the above limitations

Possible approaches

• Replace the current fine detector model by TLs detector model like Yolox which takes whole images as input and reuse current autoware_tensorrt_yolox package for inference.
• The interest TLs inside images which neccessrary for Planning module could be selected by a new node (traffic_light_selector)
• Remove pointcloud based autoware_traffic_light_occlusion_predictor when DNN TLs detector can ignore occluded TLs. To keep the update as minimum, current classifier will be re-used and the result will be combined by a new simple traffic_signal_merger node.
  The new pipeline as figure below
  

Definition of done

• Training and upload new TLs detector model
• Update autoware for new model downloads
• Implement traffic_light_detector package
• Add executor for new TLs detector model
• Add traffic_light_signal_merger
• Update tier4_perception_launcher for new pipeline
• Update autoware_documentation and perception diagram