A visual-assist wearable for the blind and visually impaired, from research concept to Versal AI Edge prototype
Industry hardware partner on an NIH-funded assistive-vision program: headset and compute-pack architecture on a Versal AI Edge module, event and RGB camera integration, board and camera-interface design, and a SLAM front-end accelerator concept.
The challenge
Research groups at two universities are building a wearable that helps blind and visually impaired people navigate and recognize their surroundings. The algorithms are the researchers’ domain. Turning them into a headset and compute pack that runs in real time, on a body, with cameras where they need to be, is ours.
What we owned
System architecture. A headset and compute pack built on an AMD Versal AI Edge system-on-module, with an event camera alongside RGB and depth sensors and an IMU per camera position. Sensor-bridge and camera-interface design, including MIPI-to-GMSL2 and FPD-Link options so far-side cameras can sit across the head from the FPGA without signal-integrity compromises.
Board design. The smart-camera PCB and schematics, connector and flex-cable planning, and the mechanical envelope for the headset.
Accelerator concept. A SLAM front-end for the programmable logic: streaming keypoint detection with a row-cache buffer, keypoint filtering and ranking, patch collation and DMA into on-chip memory, with descriptor computation mapped onto the Versal AI Engine array. Feature-descriptor requirements were compared against AI Engine capabilities to decide the hardware/software partition.
Continuity. This extends a decade of assistive-vision work: the NSF Expedition visual-prosthesis prototypes and the HoloLens-plus-POWER8 recognition system demonstrated on Capitol Hill in 2016.
Outcome
A prototype platform under active development with the research teams, engaged under a university professional-services agreement. Institution and program naming shown here with permission pending.