Skip to content

Releases: Corning-AI/embedded-linux

v0.3.0 — Real-Time Edge AI: NPU Object Detection at 11ms

09 Mar 17:44
@Corning-AI Corning-AI
v0.3.0
e23fcad

Choose a tag to compare

View all tags
v0.3.0 — Real-Time Edge AI: NPU Object Detection at 11ms Latest
Latest

Real-Time Edge AI Demo on i.MX 8M Plus

Watch the attached demo video (68 seconds) showing live NPU-accelerated object detection on the i.MX 8M Plus EVK.

What's in the video

  • Live OV5640 MIPI camera → NPU inference → HDMI output with bounding boxes
  • 11ms per frame on the 2.3 TOPS NPU (VX Delegate + TFLite INT8)
  • Multi-object detection: person, cell phone, bottle — up to 3 objects simultaneously
  • Real-time OSD showing model name, NPU latency, FPS, and object count
  • Stable 9 FPS end-to-end pipeline (camera capture → inference → overlay → display)

Performance

Model NPU CPU Speedup
MobileNet SSD v2 (80 COCO classes) 11 ms 45 ms 4×
MoveNet Lightning (17 keypoints) 13 ms 26 ms 2×

What's new in v0.3.0

  • Revamped README with demo screenshots and performance showcase
  • 5 demo frames extracted from video for the GitHub page
  • Updated repo description and topics for discoverability
  • Detection app improvements (multi-model, OSD, headless mode)

Full pipeline

OV5640 (MIPI CSI-2) → ISI DMA → GStreamer appsink → TFLite + NPU → PIL overlay → HDMI

Previous releases

  • v0.2.0: Kernel drivers, V4L2 capture, device tree, WiFi/BT
Assets 3
Loading

v0.2.0 — Kernel drivers, camera pipeline, NPU stack

09 Mar 16:26
@Corning-AI Corning-AI
v0.2.0
b22affb

Choose a tag to compare

View all tags
v0.2.0 — Kernel drivers, camera pipeline, NPU stack

What's new

Kernel modules

  • hello — Minimal loadable module demonstrating module_init/module_exit and printk
  • chardev — Character device driver with file_operations (read/write/ioctl), mutex synchronization, automatic /dev node creation
  • bme280 — I2C client driver with device tree compatible matching, i2c_smbus register access, and sysfs attributes

Userspace programs

  • V4L2 capture (drivers/v4l2-capture/) — Frame capture using the V4L2 multi-planar API, mmap zero-copy buffers, directly targets the i.MX8MP ISI
  • Camera detection app (app/camera-detect/) — TFLite + VX Delegate object detection with GStreamer camera input

Device tree

  • OV5640 overlay (dts/imx8mp-evk-ov5640.dts) — Fully annotated device tree overlay showing MIPI CSI-2 pipeline, clock providers, regulator bindings, and OF graph endpoint linking

Hardware verified

  • OV5640 MIPI CSI-2 camera: live 640x480@30fps preview on HDMI via GStreamer
  • NPU stack: galcore 6.4.11 + VX Delegate + TFLite 2.16.2
  • GPU: Weston + weston-simple-egl @ 60 FPS

Documentation

  • Rewrote README with actual repository structure and verified hardware results
  • Updated camera+NPU guide with correct /dev/video3 device path and ISI pipeline details

Target hardware

NXP i.MX 8M Plus EVK (8MPLUSLPD4-EVK), Yocto Scarthgap, kernel 6.6.52-lts

Loading