Using HailoTracker with Custom Object Detection Model (Non-GStreamer Implementation)

Sai_Raghava_Munagala · November 10, 2025, 1:29pm

Hello Hailo Community,

I’m working on a drone detection and tracking system on Raspberry Pi 5 with Hailo AI HAT, and I need guidance on implementing HailoTracker with my custom object detection model.

My Requirements:

Custom-trained object detection model (YOLOv8/v10 for drone detection)
Need to integrate HailoTracker for multi-object tracking
Want to use HailoTracker without GStreamer (prefer direct Python/C++ API)
Target: Real-time tracking at 20-30 FPS

Current Setup:

Hardware: Raspberry Pi 5 + Hailo-8L AI HAT
Detection model: Custom .hef file (already compiled and working)
Framework: Python with HailoRT

Questions:

Is there a standalone HailoTracker API that can be used independently of GStreamer elements? I've seen references to hailotracker in the GStreamer plugin, but I need to use it in a custom Python pipeline.
What tracking algorithm does HailoTracker use internally? (SORT, DeepSORT, ByteTrack, etc.) This will help me understand performance characteristics for fast-moving small objects.
Does HailoTracker support custom Re-ID features? For my use case (anti-lost drone tracking), I need appearance-based re-identification after occlusions.
Integration approach: What's the recommended way to integrate tracking with custom detection models? Should I:
    Extract tracking functionality from GStreamer plugin
    Use HailoRT + separate tracker (ByteTrack/OC-SORT)
    Build custom tracker using Hailo acceleration
Sample code/examples? Are there any examples of using HailoTracker outside of the standard GStreamer pipeline, especially for custom detection models?

omria · November 10, 2025, 1:59pm

Hey @Sai_Raghava_Munagala ,

You can check how we do the tracker in here :

For more info about the HailoTracker :

github.com/hailo-ai/tappas

docs/elements/hailo_tracker.rst

master

Hailo Tracker
==============

Overview
--------

| HailoTracker is an element which enables the user to track `HailoDetection <../write_your_own_application/hailo-objects-api.rst#hailodetection>`_ objects using the Joint-Detection-and-Embedding (JDE) Tracking algorithm. JDE is a fast and high-performance multiple-object tracking algorithm that associates between detections in consecutive frames based on size/location/movement. Associations are made by linear assignment via a Kalman Filter.
| For more information on `JDE Tracking and Kalman Filtering read here <https://github.com/Zhongdao/Towards-Realtime-MOT>`_.

Parameters
^^^^^^^^^^

The hailotracker element provides a series of properties that allow you to adjust the tracking algorithm. The most important property to set is ``class-id``\ : this determines if the tracker will track all `HailoDetection <../write_your_own_application/hailo-objects-api.rst#hailodetection>`_ objects indiscriminately of class or focus only on detections of a specific class id (the default behavior is to track across-classes). 

Hierarchy
---------

.. code-block::

   GObject

This file has been truncated. show original

Sai_Raghava_Munagala · November 11, 2025, 10:10am

I’m working on developing a custom tracker that leverages the Hailo AI HAT for acceleration. I’ve gone through the documentation at docs/elements/hailo_tracker.rst which explains the standard tracker implementation.

I’m looking to understand:

How to implement a custom tracking algorithm that utilizes the Hailo AI HAT’s processing capabilities
Whether the hailo_tracker element supports custom tracking backends or if I need to develop a custom GStreamer element
How to properly interface my custom tracker with the HAT to take advantage of hardware acceleration
Any API or SDK references for developing custom tracking solutions on the Hailo platform
Performance considerations when implementing custom tracking vs. using the built-in tracker

My goal is to implement a specialized tracking algorithm while maximizing the computational efficiency provided by the Hailo AI HAT.