StreetSense

StreetSense is a camera-based street-scene assistant prototype for blind and low-vision pedestrians. It turns a busy camera feed into short, prioritized spoken cues about crossings, pedestrian signals, moving vehicles and nearby obstacles. The system has two loops. Local computer vision detects and tracks people, vehicles and signal heads, understands road/sidewalk surfaces, and reads pedestrian signals. Background Meta AI adds scene context and an independent crossing assessment. A deterministic decision layer controls what can be spoken: crossing permission requires both a local WALK reading and a validated positive Muse Spark response. Unclear, conflicting, malformed or stale evidence cannot authorize crossing. The demo shows recorded operation on real street footage, including a case where Muse Spark says yes but the local reader is unclear and the system refuses a go cue. It also shows a nearly expired countdown, European-style walking figures, actual Segment Anything masks, and a moving-vehicle alert spoken from the selected local pipeline output. Recorded model replies and replayed rule tests are labelled. Engineering work includes country/style profiles for UK and European signals, bounded signal-head refinement, per-track temporal voting, timestamp-based expiry, generation checks for asynchronous replies, scene-reset handling, prioritized speech and segmentation caching. The current test suite has 479 passing tests and one skipped test. This is a prototype evaluated on public recordings, not a validated mobility aid or a completed field study. Next steps are accessibility-led evaluation and calibrated real-world measurements. Source, reproduction instructions, measured limitations and media credits are available in the repository.