Algorithm for detecting foreign objects on tram tracks (dRoI) based on segmentation

Weights available in HuggingFace: here.

---

Launcher UI

ui/app.py is a minimal desktop launcher: pick Camera Mode (live camera) or Video Mode (drop/select a video file), set the tram speed, and it runs algorithm/detection.py for you. Also you can test it on Image Mode (drop/select a image file).

python ui/app.py

Drag-and-drop in Video Mode requires the optional tkinterdnd2 package (pip install tkinterdnd2); without it, use the "Обзор…" file picker button instead.

---

Example:

---

Graduation Thesis (FQW/ВКР): Algorithm for detecting foreign objects on tram tracks based on their segmentation.

Seven segmentation models were researched and trained (YOLO-seg: 8n, 8s, 11s, 26s; SegFormer: B0, B3; DeepLabV3+: ResNet50), as well as three detectors: YOLOX, YOLOv8, and YOLOv5.

Based on the research results, the following were selected for the task:

• Segmentor: YOLOv11s-seg
• Detector: YOLOv8s
• Tracker: ByteTrack

The developed algorithm is available at: https://github.com/MathematicLove/tram-dynamic-roi-tracker

Useful Links

• Research paper (article): https://elibrary.ru/qfcwed
• Diploma for 1st degree laureate (Best report in ML and Intelligent Data Processing section): Certificate
• Full Thesis (FQW/ВКР): coming soon

---

Available Models (Pickle)

---

by Salimli Ayzek (Салимли Айзек): https://mathematiclove.github.io