Integrating Re-ID Model into YOLO Detection Pipeline on Hailo-RPI5

Hello,

I’m currently testing with the hailo-rpi5-examples, and my goal is to combine a person Re-ID model with the YOLO detection pipeline.

Specifically, I want to be able to continue tracking a specific object (person), even if they temporarily disappear from the camera view. Once they reappear, I want the system to reassign the same track ID to that specific person, so tracking continues seamlessly.

After thoroughly reviewing the example codes provided in hailo-rpi5-examples, I believe it’s possible to integrate the Re-ID model into the INFERENCE_PIPELINE, and use it within the TRACKER_PIPELINE to assign track IDs based on feature similarity.

If anyone has experience working on a YOLO + Re-ID project using the TAPPAS framework, I would really appreciate any advice or guidance you could share.

Thank you!

Hey @Sangjoon_Shin,

Welcome to the Hailo Community!

Yes, this is absolutely possible. I recommend starting with the following resources to understand how the pipeline components fit together:


Suggested Flow for YOLO + Re-ID

  1. Detection Pipeline

    • Run a YOLO model to detect people (class_id=1) and produce bounding boxes.
  2. Tracking Pipeline

    • Use hailotracker with keep_past_metadata=True to maintain persistent track info.
  3. Re-ID Pipeline

    • Use CROPPER_PIPELINE to extract cropped detections.
    • Run a Re-ID model on those crops in parallel.
    • Compare extracted features to an in-memory feature database for identity matching.
  4. Identity Reassignment Logic

    • On Re-ID match, assign consistent global IDs to tracks using hailotracker.
    • Enables recovery from occlusion — when someone reappears, their embedding helps reassign the correct ID.
1 Like