What is metro-navigation-public?

This project is a research prototype for a computer vision system that could help visually impaired people navigate metro stations. It watches video from a camera, using several AI models together: two object detectors trained to spot metro landmarks, warning signs, and station signage, plus a segmentation model that identifies safe walking paths and curb ramps in the image. A guidance system then combines all of this into spoken instructions, moving through different modes such as searching for a safe starting area, following a tactile path, and warning about obstacles or announcing arrival at a destination. The README is direct about what this project is and is not. It states clearly that this is a research prototype, not a certified mobility aid, and that it has not been tested for safe unsupervised use. It should not be relied on as someone's only means of navigation. The repository currently includes the code for running the system and benchmarking its components, but it does not yet include the full training pipeline, the dataset used to train the models, or the official test data, so results cannot be fully reproduced by an outside user yet. To run it, a person needs Python 3.10 or 3.11, ideally a CUDA-capable graphics card for real-time speed, though it can run on a CPU for basic testing, plus a text-to-speech engine and three separate pretrained model weight files placed in a specific folder. Once set up, the tool can process a video file or a live camera feed, optionally displaying results on screen and speaking instructions aloud, and it can save an annotated video along with a log of performance data. A separate benchmarking command measures how fast each part of the pipeline runs. The authors are honest about the system's current limits: it has only been tested on a single, relatively uncrowded route, and its reported accuracy does not prove it is safe in crowded, poorly lit, rainy, or unfamiliar stations. Turning this into something safe to deploy would require substantially more testing, safety engineering, and validation with real users.