Dynamic Real-time Adaptive Coordinates for Drone Trajectory Prediction

DRAC-Prediction is a pipeline for detecting a drone in video and predicting it's future trajectory several frames ahead. Given a detector's bounding box output, the system tracks a single target across frames, derives its center-point velocity, and feeds a short history window into a GRU-based recurrent model to forecast the drone's position at multiple future horizons (e.g. +1, +2, +4, +8, +16 frames). The system is built to be modular: any video source (webcam, file, or stream), any upstream object detector, and any trained trajectory model can be swapped independently.

BBC

NOTE: To run you need the models! Download the trained models from this Google Drive and place them in the backend/models/ directory.

Table of Contents

• Demo
• Architecture
• Dataset
• Results
  • Detection
  • Trajectory Prediction

Demo

Architecture

Video Source (webcam / file / stream)
        |
        v
  Object Detector  -->  Single-Target Tracker  -->  Trajectory Inferencer
  (per-frame bbox)      (locks onto one drone)      (GRU: history > future positions)

The trajectory model (AdaptiveTrajectoryGRU) takes a window of the last N frames and outputs predicted (x, y, vx, vy) at each of several future horizons in a single forward pass, avoiding compounding prediction error at longer horizons.

Formulation

At frame $t$, the detector + tracker produce a bounding box center position $\mathbf{p}_t = (x_t, y_t)$. Instantaneous velocity is estimated via finite differences over the frame interval $\Delta t$:

$$ \mathbf{v}_t = \frac{\mathbf{p}_t - \mathbf{p}_{t-1}}{\Delta t} $$

The model input is a history window of the last $N$ frames, where each timestep is a state vector combining position and velocity:

$$ \mathbf{s}_t = (x_t,\ y_t,\ v_{x,t},\ v_{y,t}) \in \mathbb{R}^4 $$

$$ X_t = (\mathbf{s}_{t-N+1}, \mathbf{s}_{t-N+2}, \dots, \mathbf{s}_t) \in \mathbb{R}^{N \times 4} $$

This window is passed through a GRU encoder, producing a hidden state $\mathbf{h}_t$ that summarizes recent motion:

$$ \mathbf{h}_t = \text{GRU}(X_t) $$

Rather than autoregressively rolling the model forward one step at a time (which compounds error at each step), $\mathbf{h}_t$ is passed through a set of horizon-specific output heads, each predicting the full state directly at a fixed future offset $k \in \mathcal{K} = {1, 2, 4, 8, 16}$:

$$ \hat{\mathbf{s}}_{t+k} = f_k(\mathbf{h}_t), \quad \forall k \in \mathcal{K} $$

All horizons are predicted in a single forward pass, so a bad prediction at $+1$ frame doesn't propagate into the $+16$ frame estimate.

Dataset

Training data consists of per-frame drone center positions with timestamps, in the schema:

Source format (whitespace-separated, 1 header line):
    frame no.            x            y
    1.000000 654.62950739 262.89502463
    2.000000 655.02817734 263.25773399
    ...

And for training the detection model, we used a public dataset of drones, witch can be found here.

Results

Detection

For detection we used a YOLOv11 model that was fine-tuned on a public drone dataset. The model achieved great performance on the validation set, with a mean average precision (mAP) of 0.85.

Trajectory Prediction

The trajectory model was trained on a dataset of drone trajectories that looked like the image below. The model was able to predict the future positions of the drone with a training loss of 0.2103 and a validation loss of 0.4945.