Table of Contents
⚠ Disclaimer: This entry may be incomplete, out of date, or inaccurate. It is AI-maintained on a best-effort basis. Do not rely on it as a sole source — verify claims independently using the sources listed below.
Summary
UAV Detection and Tracking is an open-source, single-camera pipeline that fine-tunes a Faster R-CNN ResNet101 object detector (from TensorFlow’s model zoo) to identify drones in video, then applies a Kalman filter to trace and visualize each UAV’s trajectory as an overlay line on the output video. It is a simpler, single-view, 2D pixel-space counterpart to this section’s stereo/multi-camera 3D projects — useful as a lightweight starting point or baseline rather than a deployable multi-object 3D localization system.
Key Facts
- Author: RobCaamano (GitHub)
- Type: Open-source software project — single-camera UAV detection + 2D trajectory tracking
- Status: Active — 36 stars, 5 forks, 1 open issue as of this review
- Detection model: Fine-tuned Faster R-CNN ResNet101 (
faster_rcnn_resnet101_v1_800x1333_coco17_gpu-8, TensorFlow Object Detection API) - Tracking: Kalman filter, single-camera, 2D pixel-space trajectory only (no stereo/3D geolocation)
- License: Not stated in the repository as of this review — confirm terms directly before reuse
How It Works
The pipeline is a standard fine-tuning + inference + visualization chain built on the TensorFlow Object Detection API: video is split into frames (vids_to_frames.py), annotations are converted to TFRecord format (XML_to_TFRecord.py), and a Faster R-CNN ResNet101 model pretrained on COCO is fine-tuned for UAV detection (model_main_tf2.py). Inference (UAV_FasterRCNN.py) produces per-frame bounding boxes, and a Kalman filter module (kalman_filter.py) smooths detection centroids into a trajectory, which is drawn as a trace line over the output video. Training-loss curves (classification loss, localization loss, learning rate) are included in the repo, along with example detection images and demo tracking videos, but no formal accuracy metrics (mAP, precision/recall) are published.
Limitations
- No license specified: Terms of reuse are unclear from the repository as reviewed — verify directly with the author before any derivative or commercial use.
- 2D, single-camera only: No stereo pairing or 3D geolocation — trajectory is a pixel-space overlay, not a geographic or metric 3D position, unlike Stereo Drone Tracker or eye_sky in this section.
- No published accuracy metrics: Training-loss curves are shown, but no detection accuracy (mAP), false-positive rate, or drone-vs-other-object discrimination performance is reported.
- Older detection architecture: Faster R-CNN ResNet101 is a heavier, slower two-stage detector relative to the single-stage YOLO-family models used in this section’s newer projects (e.g., Stereo Drone Tracker’s YOLOv11s) — likely a real-time-performance tradeoff on comparable hardware, though not benchmarked in the repo.