Revamp README: showcase edge AI demo with screenshots and performance metrics

Rewrite README to lead with the NPU object detection demo — hero screenshot,
3-image gallery, key metrics table (11ms NPU, 9 FPS, 80 COCO classes).
Add media/ with demo frames extracted from the 68s HDMI capture video.
Restructure content to highlight full-stack embedded Linux + edge AI scope.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Corning-AI

README.md

@@ -1,137 +1,163 @@
-# Embedded Linux on i.MX 8M Plus
+# Edge AI on Embedded Linux — i.MX 8M Plus
+
+> **Real-time object detection at 11ms per frame** on a 2.3 TOPS NPU, with live MIPI camera feed rendered to HDMI.
+> Full BSP bring-up, kernel drivers, WiFi/BT, and NPU-accelerated inference — from first boot to working demo.
+
+---
+
+## Live Demo — NPU-Accelerated Object Detection
+
+<p align="center">
+  <a href="https://github.com/Corning-AI/embedded-linux/releases/latest">
+    <img src="media/demo-hero.jpg" alt="Real-time edge AI: person 72% + cell phone 80% detected on i.MX8MP NPU at 11ms" width="480">
+  </a>
+</p>
+<p align="center">
+  <strong>Person (72%) + Cell Phone (80%)</strong> detected in real-time on the <strong>NPU</strong> — <a href="https://github.com/Corning-AI/embedded-linux/releases/latest">watch the full 68s demo video</a>
+</p>
+
+<table>
+<tr>
+<td width="33%" align="center">
+<img src="media/demo-detection-3.jpg" width="240"><br>
+<sub><strong>3 objects</strong> — person + phone + bottle</sub>
+</td>
+<td width="33%" align="center">
+<img src="media/demo-detection-2.jpg" width="240"><br>
+<sub><strong>NPU 11.3ms</strong> — consistent ultra-low latency</sub>
+</td>
+<td width="33%" align="center">
+<img src="media/demo-person-84.jpg" width="240"><br>
+<sub><strong>person 84%</strong> — high-confidence detection</sub>
+</td>
+</tr>
+</table>
+
+### Key Results
+
+| Metric | Value |
+|--------|-------|
+| **NPU inference latency** | **11 ms** (MobileNet SSD v2, INT8 quantized) |
+| **End-to-end pipeline** | **9 FPS** (camera → NPU → overlay → HDMI) |
+| **Detectable classes** | **80** (COCO: person, phone, bottle, laptop, book…) |
+| **NPU vs CPU speedup** | **4× faster** (11ms vs 45ms per frame) |
+| **Pose estimation** | **13 ms** (MoveNet Lightning, 17 keypoints) |
+
+### End-to-End Pipeline
 
-Heterogeneous SoC platform project built on the NXP i.MX 8M Plus EVK. Covers BSP bring-up, kernel driver development, camera/NPU pipelines, and real-time co-processing with FreeRTOS on the Cortex-M7.
+```
+OV5640 ──MIPI CSI-2──▶ ISI DMA ──▶ GStreamer appsink ──▶ TFLite + NPU ──▶ PIL overlay ──▶ HDMI
+ Camera     2-lane        /dev/video3    Python/numpy      VX Delegate       bounding      waylandsink
+ 640×480                                                    ~11ms/frame       boxes+labels
+```
+
+---
+
+## What This Project Covers
+
+This is a **full-stack embedded Linux + edge AI** project on the NXP i.MX 8M Plus EVK, covering every layer from silicon to application:
+
+| Layer | What's Done | Difficulty |
+|-------|-------------|------------|
+| **BSP** | Yocto Scarthgap custom image build (`imx-image-multimedia`) | Medium |
+| **Kernel drivers** | 3 out-of-tree modules (hello → chardev → I2C BME280) | Medium |
+| **Device tree** | Annotated camera pipeline overlay (MIPI CSI-2 → ISI) | Hard |
+| **V4L2 userspace** | C program with multi-planar mmap capture | Medium |
+| **WiFi/BT bring-up** | DTB binary patch + PCIe/UART driver loading | Hard |
+| **Camera pipeline** | GStreamer → appsink → numpy (zero-copy DMA) | Medium |
+| **NPU inference** | TFLite INT8 + VX Delegate on 2.3 TOPS NPU | Hard |
+| **Edge AI app** | Real-time detection + pose + OSD → HDMI display | Hard |
+| **Debug methodology** | 6 documented debug cases with root cause analysis | — |
+
+---
 
 ## Hardware
 
-- **SoC:** NXP i.MX 8M Plus — quad Cortex-A53 (1.8 GHz) + Cortex-M7 + 2.3 TOPS NPU
-- **Board:** 8MPLUSLPD4-EVK (6 GB LPDDR4, 32 GB eMMC)
-- **Camera:** OV5640 MIPI CSI-2 via MINISASTOCSI adapter (J12)
-- **Display:** HDMI 2.0a (J17), Weston/Wayland compositor
-- **Kernel:** 6.6.52-lts (Yocto Scarthgap)
+| Component | Spec |
+|-----------|------|
+| **SoC** | NXP i.MX 8M Plus — quad Cortex-A53 (1.8 GHz) + Cortex-M7 + **2.3 TOPS NPU** |
+| **Board** | 8MPLUSLPD4-EVK (6 GB LPDDR4, 32 GB eMMC) |
+| **Camera** | OV5640 MIPI CSI-2 via MINISASTOCSI adapter (J12) |
+| **Display** | HDMI 2.0a (J17), Weston/Wayland compositor |
+| **WiFi/BT** | AzureWave AW-CM276NF (NXP 88W8997, PCIe + UART) |
+| **Kernel** | 6.6.52-lts (Yocto Scarthgap) |
 
 ## Repository Structure
 
-```text
+```
+app/camera-detect/         Real-time edge AI app (detection + pose + demo modes)
 kernel-modules/
-├── hello/          Minimal loadable kernel module
-├── chardev/        Character device driver (/dev node, ioctl, mutex)
-└── bme280/         I2C client driver with sysfs interface
-
-drivers/
-└── v4l2-capture/   V4L2 mmap frame capture (C, multi-planar API)
-
-dts/
-└── imx8mp-evk-ov5640.dts   Annotated device tree overlay (camera pipeline)
-
-debug/
-├── 01-video-device-numbering   /dev/video0 is NOT the camera on i.MX8MP
-├── 02-camera-red-tint          ISI RGB/BGR format mismatch → color shift
-├── 03-galcore-not-in-lsmod     Built-in driver vs loadable module confusion
-├── 04-wifi-dns-resolution      WiFi connected but no DNS → resolv.conf
-├── 05-media-controller-pipeline  MC API graph setup for MIPI camera
-└── 06-weston-service-name      Service renamed between Yocto releases
-
-scripts/
-├── build-multimedia.sh      Yocto build helper for imx-image-multimedia
-└── serial_transfer.py       File transfer over debug UART (base64)
-
-docs/
-├── 01-dev-environment.md    Build host setup
-├── 02-hardware-guide.md     EVK wiring, boot config, peripheral map
-├── 03-yocto-bsp.md          Yocto BSP build (repo init → bitbake)
-├── 07-camera-npu.md         Camera + NPU object detection pipeline
-└── 08-device-tree-explained.md  Device tree walkthrough
+├── hello/                 Minimal loadable kernel module
+├── chardev/               Character device driver (/dev node, ioctl, mutex)
+└── bme280/                I2C client driver with sysfs interface
+drivers/v4l2-capture/      V4L2 mmap frame capture (C, multi-planar API)
+dts/                       Annotated device tree overlay (OV5640 → ISI pipeline)
+debug/                     6 documented hardware/driver debug cases
+scripts/                   Yocto build helper, serial file transfer tool
+docs/                      Step-by-step guides: BSP, hardware, WiFi/BT, camera, NPU
 ```
 
-## What's Working
-
-| Subsystem | Status | Details |
-| --------- | ------ | ------- |
-| Yocto BSP | Boots from SD | `imx-image-multimedia`, Scarthgap branch |
-| OV5640 camera | Live HDMI preview | MIPI CSI-2 → ISI → GStreamer → Weston |
-| NPU (VIP8000) | Driver loaded | galcore 6.4.11, VX Delegate + TFLite 2.16.2 |
-| GPU (GC7000UL) | Weston @ 60 FPS | `weston-simple-egl` verified |
-| Debug UART | Working | J23, 3rd COM port = A53 console (115200 8N1) |
-
 ## Quick Start
 
 ```bash
-# 1. Build the Yocto image (Ubuntu 22.04 host, ~2 hours first build)
+# 1. Build the Yocto image (Ubuntu 22.04 host)
 source scripts/build-multimedia.sh
 
-# 2. Flash to SD card (use Rufus DD image mode on Windows)
-#    Set SW4: OFF OFF ON ON (SD card boot)
-
-# 3. Connect: USB-C to J5 (power), micro-USB to J23 (debug UART)
-#    Serial: 3rd COM port, 115200 8N1
-
-# 4. Camera preview (on the EVK, over serial console)
-export XDG_RUNTIME_DIR=/run/user/0
-gst-launch-1.0 v4l2src device=/dev/video3 ! \
-    video/x-raw,width=640,height=480,framerate=30/1 ! \
-    videoconvert ! autovideosink
-
-# 5. Build and load the hello module (cross-compile or on-device)
-cd kernel-modules/hello
-make ARCH=arm64 CROSS_COMPILE=aarch64-poky-linux- \
-     KERNELDIR=/path/to/yocto/build/tmp/work/.../linux-imx/build
-scp hello.ko root@<EVK_IP>:/tmp/
-# On EVK:
-insmod /tmp/hello.ko && dmesg | tail -3
-```
+# 2. Flash SD card (Rufus DD mode), set SW4: OFF OFF ON ON
 
-## Kernel Modules
-
-Three out-of-tree modules demonstrating progressive driver complexity:
-
-**hello** — Module lifecycle (`init`/`exit`), `printk`, section markers (`__init`/`__exit`).
-
-**chardev** — Full character device with dynamic major allocation, `file_operations` (read/write/ioctl), `copy_to_user`/`copy_from_user`, mutex synchronization, and automatic `/dev` node creation via `class_create`/`device_create`.
-
-**bme280** — I2C client driver using the `probe`/`remove` lifecycle, device tree `compatible` matching, `i2c_smbus_*` register access, sysfs attributes, and `devm_` managed allocation.
+# 3. Connect: USB-C → J5 (power), micro-USB → J23 (debug UART, 3rd COM port, 115200)
 
-## V4L2 Capture
+# 4. Run the edge AI demo on the EVK
+export XDG_RUNTIME_DIR=/run/user/0 WAYLAND_DISPLAY=wayland-1
+python3 /opt/camera-detect/detect_camera.py --mode demo
+```
 
-Userspace C program demonstrating the V4L2 multi-planar API as used by i.MX8MP's ISI:
+## Edge AI Detection App
 
-- `VIDIOC_QUERYCAP` → capability query
-- `VIDIOC_S_FMT` → format negotiation (multi-planar)
-- `VIDIOC_REQBUFS` + `mmap()` → zero-copy DMA buffer setup
-- `VIDIOC_QBUF`/`VIDIOC_DQBUF` → streaming capture loop
+```bash
+python3 detect_camera.py                         # Object detection (MobileNet SSD v2)
+python3 detect_camera.py --mode pose              # Pose estimation (MoveNet, 17 joints)
+python3 detect_camera.py --mode demo              # Both models simultaneously
+python3 detect_camera.py --mode demo --compare    # NPU vs CPU side-by-side benchmark
+python3 detect_camera.py --no-display             # Headless mode (SSH/serial output only)
+```
 
-## Device Tree
+Features: real-time OSD (FPS, latency, object count), NMS post-processing, multi-model support, Wayland/HDMI output, headless mode.
 
-The [`dts/imx8mp-evk-ov5640.dts`](dts/imx8mp-evk-ov5640.dts) overlay is annotated to explain how the camera pipeline is described in hardware:
+## Kernel Modules
 
-```text
-OV5640 (I2C, 0x3c) → MIPI CSI-2 RX → ISI ch0 → /dev/video3
-```
+Three out-of-tree modules with progressive complexity:
 
-Covers clock providers, regulator bindings, OF graph endpoint linking, and MIPI lane configuration.
+| Module | Concepts |
+|--------|----------|
+| **hello** | `module_init`/`module_exit`, `printk`, `__init`/`__exit` section markers |
+| **chardev** | `file_operations`, `copy_to_user`/`copy_from_user`, mutex, dynamic major, `class_create`/`device_create` |
+| **bme280** | I2C client driver, device tree `compatible` matching, `i2c_smbus_*`, sysfs attributes, `devm_` managed alloc |
 
-## i.MX8MP Video Device Map
+## Debug Notes
 
-On this SoC, `/dev/video0` is **not** the camera:
+Real debugging cases encountered during bring-up — each with root cause and fix:
 
-| Device | Function |
-| ------ | -------- |
-| /dev/video0 | VPU H.264/HEVC encoder |
-| /dev/video1 | VPU decoder |
-| /dev/video2 | ISI memory-to-memory (CSC) |
-| /dev/video3 | **ISI capture (camera)** |
+| # | Issue | Root Cause |
+|---|-------|-----------|
+| 01 | `/dev/video0` is not the camera | VPU encoder registered first; ISI capture = `/dev/video3` |
+| 02 | Camera feed has red tint | ISI outputs BGR, app assumed RGB |
+| 03 | `galcore` not in `lsmod` | Built-in kernel driver, not a loadable module |
+| 04 | WiFi connected but no DNS | `resolv.conf` not populated by DHCP client |
+| 05 | Camera pipeline won't start | Media controller link setup required before streaming |
+| 06 | `weston@root` service not found | Renamed to `weston.service` in newer Yocto |
 
 ## Roadmap
 
-- [x] Yocto BSP bring-up and first boot verification
-- [x] Camera pipeline: OV5640 → ISI → GStreamer → HDMI
-- [x] NPU stack verification (galcore, VX Delegate, TFLite)
-- [x] Kernel module examples (hello, chardev, I2C driver)
-- [x] V4L2 capture program and device tree overlay
-- [ ] NPU benchmark: CPU vs NPU inference latency
-- [ ] Real-time object detection (MobileNet SSD + camera + NPU)
-- [ ] FreeRTOS on Cortex-M7 with RPMsg inter-core communication
-- [ ] End-to-end demo: camera → NPU → overlay → HDMI
+- [x] Yocto BSP build and first boot
+- [x] Camera: OV5640 → MIPI CSI-2 → ISI → GStreamer → HDMI
+- [x] NPU stack: galcore + VX Delegate + TFLite INT8
+- [x] Kernel modules: hello → chardev → I2C driver
+- [x] V4L2 capture (C) and device tree overlay
+- [x] WiFi (PCIe) + Bluetooth (UART) bring-up
+- [x] NPU benchmark: 11ms NPU vs 45ms CPU
+- [x] **Real-time edge AI: camera → NPU → overlay → HDMI**
+- [ ] FreeRTOS on Cortex-M7 + RPMsg inter-core communication
 
 ## License