Implementation:Kornia Kornia BoxMOT Tracker

Knowledge Sources	Kornia
Domains	Vision, Object_Tracking
Last Updated	2026-02-09 15:00 GMT

Overview

Wraps a detector and a multi-object tracker from the BoxMot library to provide end-to-end object tracking on video frames.

Description

The BoxMotTracker class in the Kornia contrib package combines an object detector (by default RT-DETR) with a multi-object tracker from the BoxMot library. It supports multiple tracker algorithms including BoTSORT, DeepOCSORT, OCSORT, HybridSORT, ByteTrack, StrongSORT, and ImprAssoc. The class provides an update method to feed new frames, a visualize method to render tracking trajectories, and a save method to write output images to disk. The tracker requires at least 4 frames to initialize tracking positions.

Usage

Import this module when you need to perform multi-object tracking across video frames, combining a detector with a tracking algorithm that maintains object identities across frames.

Code Reference

Source Location

• Repository: Kornia
• File: kornia/contrib/boxmot_tracker.py
• Lines: 1-182

Signature

class BoxMotTracker:
    def __init__(
        self,
        detector: Union[ObjectDetector, str] = "rtdetr_r18vd",
        tracker_model_name: str = "DeepOCSORT",
        tracker_model_weights: str = "osnet_x0_25_msmt17.pt",
        device: str = "cpu",
        fp16: bool = False,
        **kwargs: Any,
    ) -> None: ...

    def update(self, image: torch.Tensor) -> None: ...

    def visualize(self, image: torch.Tensor, show_trajectories: bool = True) -> torch.Tensor: ...

    def save(self, image: torch.Tensor, show_trajectories: bool = True,
             directory: Optional[str] = None) -> None: ...

Import

from kornia.contrib import BoxMotTracker

I/O Contract

Inputs (__init__)

Name	Type	Required	Description
detector	ObjectDetector or str	No	Detector model or name string (default: "rtdetr_r18vd")
tracker_model_name	str	No	Tracking algorithm name: "BoTSORT", "DeepOCSORT", "OCSORT", "HybridSORT", "ByteTrack", "StrongSORT", or "ImprAssoc" (default: "DeepOCSORT")
tracker_model_weights	str	No	ReID model weights filename (default: "osnet_x0_25_msmt17.pt")
device	str	No	Device for inference, e.g. "cpu" or "cuda" (default: "cpu")
fp16	bool	No	Whether to use half-precision inference (default: False)

Inputs (update)

Name	Type	Required	Description
image	torch.Tensor	Yes	Input image tensor with shape (1, 3, H, W) or (3, H, W)

Outputs

Name	Type	Description
tracking_results (from update)	numpy.ndarray	Array of shape (M, 8) with columns (x1, y1, x2, y2, id, conf, cls, ind)
visualization (from visualize)	torch.Tensor	Image tensor with shape (3, H, W) with drawn tracking results

Usage Examples

import torch
from kornia.contrib import BoxMotTracker

# Initialize tracker with default RT-DETR detector and DeepOCSORT tracker
tracker = BoxMotTracker(detector="rtdetr_r18vd", tracker_model_name="DeepOCSORT")

# Process video frames (at least 4 frames needed for initialization)
for frame_idx in range(10):
    image = torch.rand(1, 3, 640, 640)  # simulated frame
    results = tracker.update(image)

# Visualize the last frame with trajectories
vis_output = tracker.visualize(image, show_trajectories=True)

# Save output to disk
tracker.save(image, show_trajectories=True, directory="output_tracking")

Related Pages

• Principle:Kornia_Kornia_Multi_Object_Tracking