Obstacle Detector
Overview
This repository contains a fine-tuned YOLO11n object detector for obstacle and hazard detection.
The model is designed to detect obstacles in the Assist-Eye navigation pipeline.
Model Card Summary
- Base architecture: YOLO11n
- Task: object detection
- Training framework: Ultralytics
- Input size: 640 x 640
- Training strategy: transfer learning from the pretrained
yolo11n.pt checkpoint
Labels
The detector predicts the following classes:
животное, открытый люк, мусорный контейнер, яма, бордюр, забор, трещина, столб
Training Data
The training notebook expects a YOLO-style dataset layout with images and labels split into train and validation sets.
In the notebook, the dataset is loaded from unified_obstacles and the class counts are balanced with a simple oversampling pass before training.
Training Procedure
The model was trained with:
- pretrained weights:
yolo11n.pt
- epochs: 70
- patience: 7
- image size: 640
- batch size: 64
- validation during training
- early stopping based on validation improvements
After training, the best checkpoint was validated and then exported to TensorFlow Lite with INT8 quantization.
