Object Tracking

Object tracking method using a Kalman Filter and a matching process inspired by ByteTrack.

Installation

The tracking package can be installed using pip by running the following commands:

cd object-tracking
pip install -e .

Usage

Interactive tracking example

Run the following lines with Python:

from tracking import InteractiveTracker
InteractiveTracker()

Or alternatively, run:

python -m tracking.interactive_tracker

Example usage

Run the following lines with Python:

from tracking import Tracker, Detection

# Initialize the tracker
config = {}
tracker = Tracker(config)

# Get some detections
detections = [Detection(500, 100, 100, 100, "object", 0.9),
              Detection(200, 300, 200, 100, "object", 0.8)]

# Update the tracker
tracker.update(detections)

# Visualize the tracker state
tracker.show(savefig="tracker_1.png")

# Get some other detections
detections = [Detection(480, 120, 140, 120, "object", 0.95),
              Detection(200, 300, 170, 80, "object", 0.9)]

# Update the tracker again
tracker.update(detections)

# Visualize the new tracker state
tracker.show(savefig="tracker_2.png")

Method

Kalman Filter

State

The state is represented by a $\mathbb{R}^8$ vector $\mathbf x$. The state captures the position $(x, y)$ and size $(w, h)$ of the bounding box as well as their time deriatives $(\dot x, \dot y)$ and $(\dot w, \dot h)$:

$$ \mathbf{x} = (x, y, w, h, \dot x, \dot y, \dot w, \dot h)^\top $$

Dynamics

The dynamics are described by the following constant velocity model with independent gaussian noise:

where $\hat \bullet$ stands for the current mean estimate of the state.

Measurement Model

The measurements are the detections, which can be obtained via an object detector such as YOLO. The measurement model is given by:

$$ \mathbf{z} = (z_x, z_y, z_w, z_h)^\top $$

Adaptive Measurement Noise

Inspired by the NSA Kalman Filter proposed in GIAOTracker, the measurement noise should vary with the detection confidence. The following adaptive formulation is used to adapt the noise measurement:

Matching

Inspired by ByteTrack, the following three-stage matching process is employed: