Road Scene Recognition for Blind People

A computer vision system that combines YOLO segmentation with Monodepth2 distance estimation to assist visually impaired individuals in understanding road scenes. This project runs on Python 3.7 and above.

⚙️ Setup

Install the required dependencies using pip:

pip install torch==1.8.0
pip install torchvision==0.9.0
pip install ultralytics
pip install opencv-contrib-python
pip install numpy
pip install Pillow

1. Clone the repository to Monodepth2 folder :

git https://github.com/nianticlabs/monodepth2.git

Alternatively, install all dependencies at once using:

pip install -r requirements.txt

🖼️ Prediction

The system can process a single image, multiple images from a directory, or video input. Use the following command:

python test.py --input_path <path> --output_path <path> --segmentation_data <dataset>

Arguments

Parameter	Description	Type	Options
--input_path	Input file or directory path	string	- Single image: image.jpg - Image directory: ./images_folder/images - Video file: video.mp4
--output_path	Output directory for results	string	e.g., ./save_out/output
--segmentation_data	Dataset used for segmentation	string	our or mapillary

Segmentation Datasets

The project supports two different segmentation datasets with distinct class sets:

1. Mapillary Dataset Classes

• Road
• Lane Marking - Crosswalk
• Sidewalk
• Obstacle
• Car
• Person
• Traffic Light-Street
• Bike Lane
• Bicycle
• Traffic Light-Sidewalk
• Pedestrian Area

2. Our Dataset Classes

• Bike
• Bikelane
• Car
• Crosswalk
• E-Scooter
• Obstacle
• Person
• Road
• Sidewalk
• Stairs
• Traffic Light