Edge AI with benchmarks attached, not marketing adjectives.
ADAS and in-cabin features increasingly run at the edge for latency, privacy, and cost. This experiment shows inference on Coral Edge TPU and i.MX93 Ethos-U and publishes results on the same CAN-FD network the rest of the bench uses.
We fix the detection message contract in the database, simulate it on Virtual Bus while perception firmware matures, then integrate on hardware with measured latency and power.
Technical Overview
Hardware: Coral Dev Board with Edge TPU and camera, FRDM i.MX93 with Ethos-U65, Raspberry Pi orchestration, power profiling, and CAN-FD interfaces.
Detection structures come from the code generator; synthetic replay and bus budget checks run before live camera integration.
Side-by-side accelerator comparison on the same detection task with measured power per inference.
What This Demonstrates
- Live inference on Coral Edge TPU and i.MX93 Ethos-U
- Camera detection pipeline
- Detection messages on CAN-FD
- Latency and power-per-inference measurements
- A/B accelerator comparison on one task
- Simulation-first detection contract
- Instrument-cluster ADAS alert integration