Prefer not to manually configure devices or enter commands? Let AI Assistant size hardware based on your site setup, or run one-click script deployment once ready.
One Jetson Orin runs everything: the MQTT broker, the alert workbench, and the person detector (TensorRT inference, NVDEC decode). No second machine is required.
Enter the machine's address and the camera's RTSP URL; the engine is built and three containers are started.
Engine written:. Do not interrupt it.mqtt_connected true and "decode": "hw".📖 Engineering Reference: Siting constraints and on-site wiring guidelines are documented in the Engineering Wiki · Deployment Guide.
sudo apt-get install -y nvidia-container-toolkit && sudo nvidia-ctk runtime configure --runtime=docker && sudo systemctl restart docker.💡 Technical Parameters & Tuning: For Multi-stream capacity, confidence tuning, and temporal state machine, see Engineering Wiki · Measured Benchmarks.
| Symptom | Action |
|---|---|
| Deploy stops at "This preset is Jetson-only" | The target is not a JetPack machine. Use the RK3588 preset, or a Jetson. |
| Deploy stops at "The nvidia container runtime is not registered" | Install nvidia-container-toolkit as in the prerequisites, restart Docker, and deploy again. |
| Engine build fails | Read the trtexec output above the failure. Out of disk is the usual cause. |
Detector logs deserialize_cuda_engine returned None | Delete models/yolov8n_fp16.orin.engine and redeploy; the engine is rebuilt. |
| Detector exits with a hardware-decode error | Confirm the nvidia container runtime is registered, then redeploy. |
| Hub does not answer on 8090 | docker compose logs hub. A port already in use is the usual cause. |
| Detector container restarts in a loop | docker compose logs detector. An unreachable RTSP URL is the usual cause; test it in VLC from the same machine. |
| Workbench shows no device | Wait 30 s, then check that mqtt_host in config/detector.yaml is mosquitto. |
| Alerts fire twice for one person standing still | Check the camera and the network for stalls first. |
Also written into the detector config as the preview address the workbench uses
recomputerCommon defaults: recomputer (Seeed reComputer J30 / J40 series), nvidia, or ubuntu.
site-01Second segment of the MQTT topic and the name shown in the workbench device list
cam-0Appended to the topic so several cameras on one detector stay separable
H.264 only — the NVDEC path is built for H.264. Test the address in VLC first; a wrong path or wrong credentials is the most common failure.
Log in, draw one boundary, and walk across it to see the first alert.
| Symptom | Action |
|---|---|
| Login rejects the password | Run docker compose logs hub | grep -i admin to read the generated password. |
| Rule editor shows a grey canvas instead of a frame | Set preview_advertise_host in config/detector.yaml to the machine's LAN address, not 127.0.0.1. |
| Alerts have no snapshot thumbnail | Check the detector log. |
| Nothing fires when you cross the line | On the Devices page, confirm the detection rate is normal and the person is detected continuously. |
The broker, the hub and the detector are running on this machine.
http://<machine>:8090.admin with the random password the deploy step printed. If the output was truncated, run docker exec <hub container> cat /data/initial-password.txt on the machine; line 2 is the password. You must change it on first login.hw.forward, backward or any.Workbench → Video wall. Choose 1, 2, 4, 6 or 9 tiles, or Auto. Each tile shows the detection boxes plus that stream's zone and tripwire. Press F for fullscreen and Esc to exit. A tile whose device is offline or whose stream is reconnecting turns grey and shows the reason.
At the 1, 2 and 4 layouts each tile has a confidence slider. The change applies from the next frame and survives a restart. Lower is more sensitive with more false alarms; higher is more conservative. If the console says the result is unknown, reload the Devices page and read the value the detector reports.
Video wall → Add camera. Pick the detector, paste the RTSP address and name it; when the button completes, the tile is live. After adding one, check the detector's CPU figure on the Devices page before adding the next. The trash icon on a tile removes a camera; its alerts are kept.
Each alert can be acknowledged or marked a false positive, with undo, batch mode and CSV export. The interface is available in English and Chinese.
| Topic | Contents |
|---|---|
sensecraft/security/<device_id>/detections/<stream_id> | per-frame person boxes with track ids |
sensecraft/security/<device_id>/status | online status and decoder in use |
sensecraft/security/<device_id>/events/<stream_id> | alert events judged by the hub |
In this preset port 1883 accepts local connections only. To let another machine subscribe, change the mosquitto port mapping in the compose file from 127.0.0.1:1883:1883 to 1883:1883, add a password_file to config/mosquitto.conf, and set matching credentials in config/detector.yaml. Opening the port without a password lets any device on the network publish forged alerts.
Add it from the video wall, run another detector container on this machine with its own device_id, stream_id and preview_port, or deploy a separate board with the RK3588 preset.
One RK3588 board runs everything: the MQTT broker, the alert workbench, and the person detector (NPU inference, on-board hardware decode). No second machine is required.
Enter the board address and your camera URL; three containers are started on the board.
"decode": "hw".📖 Engineering Reference: For site survey constraints and wiring details, see Engineering Wiki · Deployment Guide.
rknpu2 runtime installed (/usr/lib/librknnrt.so exists), and rknn_toolkit_lite2 importable by the board's python3./dev/mpp_service present; the detector does not run on CPU decode.| Symptom | Action |
|---|---|
| Deploy stops at "rknnlite is not importable" | Run pip3 install rknn_toolkit_lite2 on the board, under the system python3 rather than a venv. |
| Detector exits complaining about the decoder | Install gstreamer1.0-rockchip1 and gstreamer1.0-plugins-bad on the board and redeploy. |
| Deploy succeeds but the detector is still an older build | docker pull the image on the board, then redeploy. |
| Model fails to load with a version error | Run strings /usr/lib/librknnrt.so | grep 'librknnrt version' to read the actual version; the shipped model is built for 2.3.2. |
W Query dynamic range failed on every start | No action needed. |
| Hub does not list the detector | Wait 30 s; confirm mqtt_host in config/detector.yaml is mosquitto. |
| Hub does not answer on 8090 | docker compose logs hub. A port already in use is the usual cause. |
Where the workbench will be served — http://<this address>:8090 — and also the preview address written into the detector config
radxaDepends on the board image — commonly radxa, cat, linaro or ubuntu.
rk3588-01Second segment of the MQTT topic and the name shown in the workbench device list
cam-0Appended to the topic so several cameras on one board stay separable
H.264 only on this board — the hardware decoder path is built for H.264
Open the workbench and confirm hardware decode.
| Symptom | Action |
|---|---|
Device page shows decode sw | Turn require_hw_decode on in config/detector.yaml and check MPP hardware decode. |
| Duplicate alerts for one motionless person | Compare the detection rate on the Devices page with the camera's frame rate. |
| Rule canvas is grey | Set preview_advertise_host to the board's LAN address. |
The workbench is at http://<board>:8090.
admin with the random password the deploy step printed (also in /data/initial-password.txt inside the hub container), and set a new password when prompted.hw.The board uses the int8 model. If false alarms are frequent, raise conf_threshold first. For wide-angle overhead sites where people appear far away, verify detection on footage from that site first.
mqtt_host in each board's config/detector.yaml to that machine's address.One board with a Hailo-8 runs everything: the MQTT broker, the alert workbench, and the person detector. No second machine is required.
decode: "sw". Watch CPU load when adding cameras.Enter the board address and your camera URL; three containers are started on the board.
"decode": "sw".📖 Engineering Reference: For site survey constraints and wiring details, see Engineering Wiki · Deployment Guide.
/dev/hailo0 exists) and the matching hailort package installed (/usr/lib/libhailort.so exists).hailort-*-cp311-*_aarch64.whl) matching the driver version under your home directory, downloaded from the Hailo developer zone.| Symptom | Action |
|---|---|
| Deploy stops at "/dev/hailo0 is already held by" | Stop the program using the accelerator (commonly another vision container) and redeploy. |
Detector exits with HAILO_OUT_OF_PHYSICAL_DEVICES(74) | Same as above. |
| Deploy succeeds but the detector is still an older build | docker pull the image on the board, then redeploy. |
| Deploy stops at "No hailort cp311 wheel found" | Download the wheel matching your driver version and place it under your home directory. |
| Detector exits with 139 or 135 | Pull the latest image and redeploy. |
Devices page shows decode: "sw" | Expected on this board. |
| Hub does not list the detector | Wait 30 s; confirm mqtt_host in config/detector.yaml is mosquitto. |
| Hub does not answer on 8090 | docker compose logs hub. A port already in use is the usual cause. |
Where the workbench will be served — http://<this address>:8090 — and also the preview address written into the detector config
piDepends on the board image — commonly pi or recomputer on Raspberry Pi OS.
hailo-01Second segment of the MQTT topic and the name shown in the workbench device list
cam-0Appended to the topic so several cameras on one board stay separable
Decoded in software on this board, which costs CPU — budget for it before adding a second camera
Open the workbench and draw your first rule.
| Symptom | Action |
|---|---|
| Live frame does not load in the rule editor | Set preview_advertise_host in config/detector.yaml to the board's LAN address rather than 127.0.0.1. |
| Alerts arrive without snapshots | If the broker restarted, wait 30 s for the detector to reconnect. |
The workbench is at http://<board>:8090.
Log in as admin with the random password the deploy step printed; you must change it on first login. If the output was truncated, run docker exec edge_security_hailo-hub-1 cat /data/initial-password.txt on the board; line 2 is the password.
Open Rules, pick this board's camera, and draw on the live frame: a polygon is a restricted zone, a line is a tripwire whose arrow points forward. Rules survive a resolution change; redraw them if the camera moves.
Workbench → Video wall. Choose 1, 2, 4, 6 or 9 tiles, or Auto. Each tile shows the detection boxes plus that stream's zone and tripwire. Press F for fullscreen and Esc to exit. A tile whose device is offline or whose stream is reconnecting turns grey and shows the reason.
At the 1, 2 and 4 layouts each tile has a confidence slider. The change applies from the next frame and survives a restart. Lower is more sensitive with more false alarms; higher is more conservative. If the console says the result is unknown, reload the Devices page and read the value the detector reports.
Video wall → Add camera. Pick the detector, paste the RTSP address and name it; when the button completes, the tile is live. The trash icon on a tile removes a camera; its alerts are kept.
Detections go to sensecraft/security/<device_id>/detections/<stream_id>, alert events to .../events/<stream_id>, and online status to .../status. Box coordinates are normalized against the original frame.
Each stream adds CPU decode load. Add one, check the detector's CPU figure on the Devices page, and add the next only if there is room.
Use this only when several detector devices are already running and should share one alert list. The Jetson, RK3588 and Hailo presets each include a hub and do not need this preset.
This preset runs the broker and the alert workbench on an always-on machine, with no detector. After deploying, change mqtt_host in each detector's config/detector.yaml from mosquitto to this machine's address and restart the detector.
Enter this machine's address; the broker and the hub are started.
📖 Engineering Reference: For site survey constraints and wiring details, see Engineering Wiki · Deployment Guide.
| Symptom | Action |
|---|---|
| Hub does not answer on 8090 | docker compose logs hub. A port conflict is the usual cause. |
| Detectors connect and drop repeatedly | Run only one hub per broker. |
| Alerts stop after the first few | Check handler_errors in /api/health; if it keeps rising, read the hub log. |
Detectors will be pointed at this address on port 1883
ubuntuLog in and change the password; the device list stays empty until a detector connects.
| Symptom | Action |
|---|---|
| Detector shows online then offline in a loop | Give each detector a distinct name. |
| Alerts appear with no thumbnail | The detector did not answer the snapshot request; the alert is still recorded. Check the detector log. |
| Device clocks differ between sites | No action needed; rule timing and ordering use the hub's clock. |
The hub is running and waiting for detectors on port 1883.
http://<machine>:8090.admin with the random password the deploy step printed (also in /data/initial-password.txt inside the hub container), and set a new password when prompted.On each detector device, edit config/detector.yaml next to its compose file, set mqtt_host to this machine's address, and run docker compose up -d detector.
The alert database, snapshots and rule configuration live in the data/ directory next to the compose file. Back up that directory.
Jetson, RK3588 and Hailo detectors can connect today; a reCamera detection node is not available yet.