Trajectory Forecast

😍😍😍 You can use any Ultralytics-supported model here. 😍😍😍

Trajectory Forecast is a lightweight, modular extension built on top of Ultralytics YOLO that enables real-time multi-object tracking with future motion prediction. It combines detection, tracking, motion history modeling, and velocity-based forecasting into a unified pipeline that can be used both as a command-line tool and as a Python library. The system is designed for practical computer vision applications such as traffic analytics, surveillance systems, robotics pipelines, and edge AI deployments.

• Installation
• Usage
  • CLI
  • Python
• Forecasting methodology
• Project structure
• Contribute

trajectory-forecasting-demo-github.mp4

Installation

pip install trajectory-forecast

Usage

CLI

Run tracking and forecasting on a video.

trajectory-forecast --model yolo26n.pt \
                    --source "https://tinyurl.com/2f3yrppv" \
                    --output result.mp4 --show --save

If you want to adjust tracking and forecasting configuration, create a config.yaml in the directory and paste the mentioned content:

# object detection confidence threshold
conf: 0.5

# tracker selection, i.e., "botsort.yaml" | "bytetrack.yaml"
# "ocsort.yaml", "deepocsort.yaml", "fasttrack.yaml", "tracktrack.yaml"                
tracker: "bytetrack.yaml"

# classes for object detection
classes: [2, 3, 5]

# store tracking history for the number of frames        
history: 40

# minimum tracking history to start calculating forecasting                 
min_points: 8

# total steps for forecasting; larger values extend the prediction horizon.             
forecast_steps: 35

# minimum speed (px/sec) before a forecast is drawn; filters out standing objects.
min_speed: 1.0

# Kalman motion flexibility; higher reacts faster, lower is smoother.
process_noise: 1.0

# Kalman detection trust; higher smooths harder (more noise assumed).
measurement_noise: 10.0

# Forecast point color (B, G, R)           
forecast_color: [255, 0, 0]

# Drawing sizes below are optional. If left out, they auto-scale to the video
# resolution. Set any of them to override.
# line_thickness: 2       # tracking box + trail thickness
# forecast_thickness: 2   # forecast line thickness
# forecast_radius: 6      # forecast marker dot radius
# font_scale: 1.2         # label text size
# font_thickness: 3       # label text thickness
# padding: 8              # label box padding

After that, you can run the code using the command mentioned below.

trajectory-forecast --model yolo26n.pt \
                    --source "https://tinyurl.com/2f3yrppv" \
                    --config "path/to/config.yaml"

Python

from tf import run_inference
from tf.config import ForecastConfig

config = ForecastConfig(
              conf=0.5,
              forecast_steps=50,
              measurement_noise=10.0,
              classes=[0, 2, 5, 6, 7]
            )

run_inference(
          model_path="yolo26s.pt",
          source="https://tinyurl.com/2f3yrppv",
          output_path="output.mp4",
          config=config
        )

Forecasting methodology

Each tracked object is smoothed with a constant-velocity Kalman filter:

• The filter keeps a running estimate of position and velocity for every track.
• On each frame it predicts the next state, then corrects it with the new detection.
• Future positions are forecast by rolling that motion model forward forecast_steps frames.
• Objects slower than min_speed are skipped so standing targets don't get a forecast.

The earlier version estimated velocity by differencing adjacent frames ((p[i] - p[i-1]) / dt), which divides a tiny per-frame delta by dt = 1/fps and so amplifies detection noise by a factor of fps — the main source of forecast jitter. The Kalman filter instead weighs each noisy detection against the predicted motion, so the estimated velocity, and therefore the forecast, stays steady. On a straight, constant-velocity track with noisy detections this cut the frame-to-frame movement of the forecast endpoint by roughly 30×.

Two knobs control the smoothing: measurement_noise (how much detections are trusted; higher smooths harder) and process_noise (how quickly the motion is allowed to change; higher reacts faster). The filter also keeps predicting through short detection gaps, which helps during brief occlusions.

Project structure

tf/
│
├── config.py        # Configuration and resolution-based auto-scaling
├── drawing.py       # Visualization utilities
├── forecasting.py   # Kalman filter and forecasting
├── tracker.py       # Per-track filter and history management
├── inference.py     # Core pipeline
└── cli.py           # Command-line interface
└── utils.py         # For downloading assets from GitHub.

Contributing

The contributions are always welcome. If you would like to extend the forecasting models or improve tracking integration, please open an issue or submit a pull request.