Skip to main content

ai-camera-viewer

:::warning [UNTESTED] on hardware — v0.9 paper-correct Builds clean on native_sim/native/64 against the full <alp/camera.h> / <alp/inference.h> surfaces (both ship today; Ethos-U dispatch runs through alp_inference_open(...) the same as any other backend). Still missing: a real Vela-compiled person_detect.tflite model and AEN HiL bench validation. :::

Headline edge-AI demo: an E1M-AEN module captures camera frames, runs a person-detect model on the on-die Ethos-U NPU, and renders the preview + bounding boxes on a 240×240 LVGL display.

Source: examples/camera-vision/ai-camera-viewer/.

Pipeline​

OV5640 → <alp/camera.h> → <alp/inference.h> → post-process → LVGL
(SCCB+MIPI)   RGB565 frame     person-detect          bbox list +
                               (uint8 quantised)      per-frame latency

What is still a placeholder​

  • No model is checked in — a Vela-compiled person_detect.tflite is still required.
  • The camera returns ALP_ERR_NOSUPPORT on native_sim; there is no emul controller for it.
  • The bounding-box decode post-process is still a TODO.

Display bring-up​

The viewer opens its panel via alp_display_open() and binds LVGL with alp_gui_lvgl_attach() before starting the inference thread, with CONFIG_LV_Z_AUTO_INIT=n in prj.conf. See lvgl-widgets-demo for the reference sequence.

board.yaml​

libraries:
  - name: lvgl
    cores: [m55_hp]
  - name: tflite-micro
    cores: [m55_hp]

som:
  sku: E1M-AEN801

preset: e1m-evk

pins:
  - { e1m: E1M_I2C0, macro: EVK_I2C_BUS_SENSORS, doc: "Shared sensor + IO-expander + INA236 bus" }
  - { e1m: E1M_SPI1, macro: EVK_SPI_BUS_ARDUINO, doc: "Arduino UNO header SPI" }

cores:
  a32_cluster:
    os: "off"
  m55_hp:
    app: ./src
    inference:
      default_arena_kib: 256
    peripherals:
      - i2c                  # Camera SCCB.
      - spi                  # Display.
      - gpio                 # Camera ancillaries + display D/C#.

chips:
  - ov5640                   # 5 MP camera.
  - st7789                   # 240x320 TFT.

diagnostics:
  log_level: info

The inference backend is driven by the SoM preset's capabilities: block, not by board.yaml — arena_kib is the only knob. Swapping som.sku compares E4 / E6 / E8 dispatch without touching the app source.

See also​

Questions about this page? Discuss in Community Forum