A prototype cane now runs real-time obstacle detection entirely offline, for less than the price of a pair of sneakers.
A team of researchers built the device by pairing a standard camera with a time-of-flight distance sensor, both run through a Raspberry Pi Zero 2W board that costs a fraction of what typical assistive tech hardware runs. An INT8-quantized SSD MobileNet V1 model processes the video feed to spot obstacles, while the distance sensor catches hazards a swinging cane tip would miss, like curbs, overhangs, or low branches. The system splits sensing, inference, and haptic feedback into separate processes so a lag in one part does not stall the whole cane. In testing, it hit a macro F1 score of 0.82, averaged 330 milliseconds of response time, and drew just 2.8 watts, light enough to run on a small battery.
The pitch is cost and independence. At $88 and fully offline, the cane needs no data plan, no server, and no subscription, which matters for the roughly 2.2 billion people worldwide living with visual impairment, many in places where cloud connectivity and pricier assistive hardware are not realistic options. It is also a rare case of edge AI built for accessibility rather than convenience or novelty.
A 12 person usability study with a System Usability Scale score of 78.5 is a promising start, not proof: that is a small sample, and 330 milliseconds of latency is plenty of time to still clip an obstacle at a normal walking pace. Worth watching if it survives contact with real sidewalks, not just a lab.