Changelog

Changelog

1.5.0

• Power-cut safe SD card. On boards that boot from a microSD card the system partition is now read-only from power-on, so pulling the power can no longer damage the operating system. Recordings and event pictures get their own partition in the part of the card the Luckfox image left unused (it used only about 7 GB of a 32 GB card). That partition is checked and repaired at every start. The installer sets this up by itself and restarts once. System > Maintenance > Power-cut protection switches it off and on.
• Less wear on the card: with recording off, the card now sees no writes at all. Settings are written only when you save them, and logs stay in RAM.
• Cards set up by an earlier test build, where the system partition fills the whole card, keep working as before. The Storage page explains how to re-flash for the new layout (docs/INSTALL.md, "Power-cut safe SD card").
• Leaner and safer: the vendor time daemon (PicoCam keeps the clock itself) and the telnet server (root login in plain text) are switched off, which frees about 1 MB of RAM. uninstall.sh puts everything back.

1.4.1

• Fixes the board freezing with 1.4.0 (no blinking light, "USB device not recognized" on the PC). The RGA on the RV1103/RV1106 has no MMU and can only read pictures in media memory; 1.4.0 handed it the AI's copy of the frame in ordinary memory, the kernel driver failed and damaged kernel memory, and the board locked up shortly after start. The AI picture now goes into a media-memory buffer the RGA can read, so RGA preprocessing works and takes about 1 ms (AI thread CPU about 2% instead of 12% on the CPU path). When no media memory is free for it, the AI prepares the picture on the CPU, and AI & Analytics says why.
• Installer: no longer reports "0 MB free" and stops when the storage name is long enough for df to wrap its line. The deploy scripts and doctor.sh read free space the same way.
• Picture right away: RTSP players, NVRs and the web live view now get their first picture in about 0.1 s instead of waiting up to 2 s for the next keyframe. The sub stream's encoder refused the immediate keyframe request and the error went unnoticed.
• Less CPU per frame: the RGA also copies the AI's frame out of the capture buffer (checked once, CPU fallback), and the exposure statistics take a quarter of the time they did. The capture thread now needs about 3.5 ms of its 66 ms per frame.
• Dashboard: NPU load is shown again (it was read from a place these boards don't have), with new tiles for the RGA, PicoCam's own CPU share, the CPU clock and the capture thread. The load average is explained: about ten idle Rockchip driver threads always count towards it, so it is not CPU use.
• Time sync without DNS: when the camera cannot resolve the NTP server, the log says so once instead of every 2 minutes, and retries slow down to every 30 minutes.

1.4.0

• No more green picture or dead exposure after a settings change. Saving Video & Audio settings, changing an AI model or a scheduled AI profile restarts the video; the camera's colour and exposure engine now restarts with it in the order the chip needs, so the picture comes back right instead of green with frozen exposure until a reboot.
• Self-healing picture: if the picture still turns green, exposure stops reacting, or the image processor stops answering, the camera restarts its video pipeline by itself (a few seconds' break, with growing pauses if it keeps happening) and records it in the audit log. Image > status shows the picture health. A video pipeline that fails to start no longer stays dead: the watchdog resets the board.
• Steadier Smart Exposure: corrections are judged only on pictures taken after the previous one took effect, a passing car or headlight no longer moves the whole scene, brightness is measured more finely, and the exposure boost no longer builds up at night (mornings no longer start over-exposed). Auto anti-flicker also switches off in bright daylight.
• Manual white balance no longer drifts with every slider move: the sliders scale the colours auto white balance had settled on.
• Less RTSP delay: a viewer that falls behind skips to a fresh picture once about 0.4 s is waiting (counting what the network stack holds too) instead of building up 2-3 s. For the lowest delay also set the player's own buffer low (VLC: Network caching 200 ms).
• Faster AI on the same NPU: the picture is scaled and converted for the model by the RGA 2D engine instead of the CPU, checked against the CPU result on first use, with automatic CPU fallback (ai.preprocess: auto or cpu). AI & Analytics shows the time spent preparing the picture and on the NPU.
• On-screen text: one switch turns off the timestamp and camera name together. Turning "Burn AI boxes into the sub stream" on or off no longer restarts the video, and the web live view no longer draws a second set of boxes on a sub stream that has them burned in.
• Sign-in: you are no longer logged out when the camera sets its clock (from NTP or the browser) after a reboot.

1.3.0

• EdgeAI server: an API for your own application. Every face and vehicle detection comes with its pictures (whole frame, cropped face, cropped plate, cropped vehicle), the person's name once saved, and the plate number with its validity and confidence. People, vehicles (one per plate, with owner, label and an allow/block flag), reports, per-camera timelines, CSV export and webhooks are available over REST. The full reference is in docs/API.md, and the server serves an OpenAPI spec at /api/v1/openapi.json.
• Alerts: rules send a message to Telegram, ntfy, email or any web address when a blocked plate or an unknown face is seen, at certain hours, or when a camera goes offline, runs hot or the server disk fills up. Each rule has a cooldown so one visitor is one alert.
• Housekeeping: automatic cleanup by age and by disk size (named people and labelled plates are kept longer), a backup download, an audit log of changes, and a call that erases a person or a plate completely.
• Safer access: several API keys with read-only or admin rights and an expiry, a sign-in page for the dashboard, expiring signed picture links, and a limit on wrong-key attempts. The server still starts open unless you give it -api-key.
• Firmware: the heartbeat to the EdgeAI server also reports temperature, CPU and NPU load, which the new health alerts use. Nothing changes on the camera when EdgeAI is off.

1.2.3

• Docs: the screenshots, banner and social preview show a real street photo instead of a drawn scene, so the AI boxes and plate reads look the way they do on a real camera. The photo is a Pexels stock image (credited in NOTICE.md), not PicoCam footage; its number plates are covered with fictional ones. No firmware or web UI changes.

1.2.2

• Safer downloads: the deploy scripts verify the download against the release's SHA256SUMS. Releases carry a build provenance attestation (gh attestation verify picocam.tar.gz --repo anand34577/luckfox-picocam).
• Small flash: the deploy scripts pick where to unpack by free space (/userdata, else RAM, else a clear message) instead of failing half way. The log kept in flash is now capped (30 lines per 5 minutes) and the known-good settings copy is only rewritten when it changed.
• Tested end to end on a fake board in CI: deploy.sh through a fake ssh into the installer, including the low-space fallback and the models-only path. deploy.ps1 was checked against a fake ssh on Windows. Unit tests for the board detection and sensor limits.
• Project hygiene: bug and feature templates, pull request checklist, Dependabot for Actions and npm, private vulnerability reporting, .editorconfig, current Action versions, and a CI check that VERSION, package.json and this changelog agree.
• Docs: screenshots of the first-run screens, and the password rule (8+ characters, letters and digits).

1.2.1

• Documentation rewritten for first-time users: a step-by-step install and flashing guide (Windows, Linux, macOS notes, USB driver, SPI NAND / eMMC / SD), a user guide with pictures of every page, an FAQ, a troubleshooting guide, and pages on connecting other software, AI models and boards. The same pages are published to the project wiki.
• Releases are built by GitHub Actions. Pushing a version tag builds the firmware and web UI from source at pinned SDK revisions, runs the tests and publishes picocam.tar.gz, picocam-lite.tar.gz, picocam-models.tar.gz and SHA256SUMS.
• Easier install: the deploy scripts need one password prompt, ignore stale SSH host keys, check the download, and explain every failure in plain words. deploy.sh --models / deploy.ps1 -Models add the ready-made AI models to a lite install.
• doctor.sh on the board: a self-check that names the likely cause of a problem, plus reset-password and factory-reset recovery commands.
• Installer fixes found by simulating it on a fake board: installing the lite package no longer fails; the patched vendor launcher is syntax-checked before it replaces the original; the Ubuntu image is refused with an explanation; a missing device-tree file no longer aborts it.
• UI: the phone header no longer squeezes the page title; System > Overview shows memory and boot flash; a demo mode (npm run mock) shows the web UI without a board.
• Demo screenshots, banner and social preview are generated from made-up data; no real footage is in the repository.

1.2.0

• Boards: run-time detection of chip, RAM, boot flash, Ethernet/Wi-Fi and camera module; one package for the whole Pico family (docs/BOARDS.md). Resolution limits and choices follow the sensor, the SD slot is found via sysfs, media-memory sizes follow the RAM.
• SPI NAND: event pictures stay in RAM without an SD card (opt-in for flash); an update checks free space and can drop the rollback copy; the installer sizes the install to /oem and there is a -lite package without models.
• Network: the USB link address (default 172.32.0.93) can now be changed in the UI, with the 60 s confirm-or-revert safety.
• Installer: preflight checks, read-only /etc fallback (start from the launcher hook), no scp needed, deploy.sh for Linux/macOS/WSL, deploy.ps1 downloads releases; both stream the package over SSH.
• UI fixes: Reconnect in live view no longer loops forever; the status badges no longer cover the camera's time overlay.
• A wrong-chip .rknn model now produces an explanatory error; convert.py --platform is validated.
• Repo prepared for open source: MIT licence, CI, docs; real-world sample footage removed from the tree.

1.1.0 and earlier

See the git history.