Safer Cycling with an Edge TPU watching your back
summary
Build uses a COCO‑trained MobileNet SSD v1 exported to TF Lite, 8-bit quantized TF Lite, then compiled for an Edge TPU USB accelerator attached to a Raspberry Pi 4 USB3. Prototype shows an LED handlebar that lights red/green based on detected buses, cars, trucks; model latency drops from ~300–400 ms on Pi CPU to ~40–50 ms with the Edge TPU and USB3, detection works on toy cars but accuracy is limited by the pretrained COCO labels and toy-scale inputs.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 4:31London cycling dangers
- 9:02Rearview detection design
- 16:04Model latency optimization
- 19:34Neural network precision
- 23:27Edge TPU hardware
- 26:18Hardware challenges
- 28:06Live demo
- 31:04Results and future plans
- 32:33Q&A
related talks
47mUnlocking Data-Driven Apps with TensorFlow
59mBuilding models for TPU devices
50mML and AI at the edge with Google Coral, EdgeTPU and Cloud Platform
27mMachine Learning on the edge using TensorFlow Lite
55mUsing TensorFlow for Performing Deep Learning at the Edge
45mUsing Edge Machine Learning to save power and development time
55mEdgeAI - AI in the Palm of Your Hand
17mBuild a smart IoT device with TensorFlow Lite and Raspberry Pi