Pre-built AI models now download from cloud cache for faster startup, pipeline branches without source are ignored, license plate recognition is improved, NVIDIA-related crashes and memory leaks are mitigated, high CPU usage bug is fixed (redeploy required)

New

• Pre-built AI model engines are now downloaded for Marketplace models from a cloud cache when available. This allows to skip "Optimizing AI model" step when starting a deployment on a gateway for the first time, which saves 15-20 minutes. Note: Pre-built Model engines are only available for common hardware configurations, so some hardware might still see longer start times.
• Pipeline Nodes that don't have a path to the source node are now ignored. This allows ignoring the whole branches of the pipeline. Previously, only Function nodes could be left unconnected.
• Support for improved License plate recognition model which is faster and more robust for side angle views of license plates.

Fixes

• Deployment runner process crashes that started happening in 1.39.12 due to NVIDIA driver issues are now mitigated with the side effect of adding some delay to starting/stopping the deployment. This is a workaround till we upgrade Deepstream to 7.1 and NVIDIA patches the DGPU drivers.
• Runners are now limited to 500 deployment starts. For deployments with a file source, the runner process stops accepting new deployments, and a new runner is started. For deployments with a live stream, the runner process is terminated and deployments are restarted. This helps to mitigate the memory leak in NVIDIA object tracker.
• Gateway version 1.39.12 introduced an issue with higher than normal CPU usage for existing deployments. That is now resolved with 1.41.33, but requires that you re-deploy (not just restart) those pipelines (Deployment detail view -> Configuration tab -> Edit configuration -> Deploy).

Known Issues

• When multiple deployments are running, logs from one deployment (in the Deployment detail view -> Logs tab) may appear in another deployment's view. We are working on a fix which will be rolled out in a patch release shortly.

Significant GPU memory savings enabling ~2x increase in number of streams processable on certain devices

New

• This update introduces a new deployment architecture that results in significant GPU memory savings per deployment. This enables you to run close to 2x the previous number of deployments for certain models on the devices that were previously memory constrained, with no change for devices/models that are GPU compute constrained:

  Device Type	Light Models	Medium Models	Heavy Models
  Jetson Orin Nano 8GB	Upto 2x more deployments	Similar capacity as v1.38.26	Similar capacity as v1.38.26
  Jetson Orin NX 8GB	Upto 1.5x more deployments	Similar capacity as v1.38.26	Similar capacity as v1.38.26
  Jetson Orin NX 16GB	Upto 1.5x more deployments	Similar capacity as v1.38.26	Similar capacity as v1.38.26
  Jetson Orin AGX 32/64GB	Similar capacity as v1.38.26	Similar capacity as v1.38.26	Similar capacity as v1.38.26
  Nvidia DGPU systems	Upto 2x more deployments	Upto 2x more deployments	Upto 2x more deployments

Updated

• Fix typo in logrotate conf which lead to some logs growing infinitely
• System metrics reporting (CPU usage and such) changed from 15 mins to 1 minute
• Fix for inbound webhooks not being passed along to Receive HTTP Webhook node.

Network route management, local recordings view and upload files for local processing.

New

• Adds Network route management, ping check and port connectivity check capabilities. See Settings -> Network.

• Add Local Recordings view

  

• Uploaded Static media files (Settings -> Media Server) can now be added to Lumeo container as a local file so you can process files without uploading them to Lumeo cloud.

  

Memory improvements during model conversion, and Save clip/snapshots to S3 now support templateable prefixes.

New

• Save Clip and Save Snapshot nodes: Uploading to customer-owned S3 buckets now supports generating the file path from a template. Check node documentation for details.
• Memory utilized during first-time model conversions is now limited, resulting in huge GPU memory savings during the process. The conversion time remains about the same.

Fixes

• Install missing build-essential package on gateway containers. Lack of this package prevented certain nodes from working.
• Increased default shared memory available within Lumeo Gateway containers to 1GB, setting stage for allowing for larger models to work.
• Model conversions are now running sequentially to avoid out of memory issues on memory constrained devices like Jetsons.

New gateway installer CLI and model conversion resource starvation fix.

New

1. Lumeo Gateway installer and manager CLI now has an improved UX, and options to install the Lumeo Gateway web interface. Learn more at Lumeo Gateway Installer.
   After running the installer once, you can now invoke it with lumeo-gateway-manager command.

2. Lumeo Web Gateway Manager v 3.0.0 feature enhancements. Learn more at Lumeo-Ready Gateways

   • Adds support for SSH, using system local credentials
   • Import cameras as streams from Genetec and Hanwha WAVE

Fixes

• When multiple deployments are started, model conversions now run sequentially to avoid running out of system memory.

Nodes gain object highlighting, runtime disable properties for alert triggers, and object attribute enhancements for flow counters. Gateway v1.34.8 addresses ONVIF caching slowdowns, adjusts monitoring intervals, and reduces load on slow VMSes with improved scheduling.

Nodes

• Display Watermark : support highlighting specified objects
• Trigger Email alert, Trigger Genetec Alarm, Trigger Immix Alarm. Trigger Milestone Event, Trigger Hanwha Wave Event : Support disabling them with a node property at runtime
• Object Flow Counter : Add crossing as obj attribute. Support publishing via Publish to Lumeo Dashboard node.

Gateway

v1.34.8 (Nov 18)

• Caching of digest auth headers is disabled again in ONVIF client because it caused slowdowns in Milestone ONVIF bridge
• Previously, if a single camera monitoring run took more time than monitoring period, then the next run started immediately. Now lumeod always has some delay between runs to reduce the load on slow VMSes.
• Changed camera monitoring period from 2 to 5 minutes, to reduce load on slow VMSes

New nodes for Hanwha Wave events, Anthropic Claude lookups, and attribute tracking enhance functionality. Updates include deployment IDs in webhooks, proximity detection improvements, and better CPU utilization for generic detectors. Platform gains bulk actions and thumbnails for search, label application from search details, and enhanced object label selection. Gateway v1.34.7 introduces ONNX improvements for YOLOv5/v8 models, snapshot thumbnails, and ONVIF auth optimizations.

Nodes

• 🍀 Trigger Hanwha Wave Event : Trigger a bookmark / event in Hanwha Wave VMS
• 🍀 Anthropic Claude Lookup : Lookup objects or scenes using Anthropic Claude
• 🍀 Track Attributes : Track object attributes and persist them over time
• Send HTTP Webhook : Add deployment id to webhook
• Custom Function Repo : Add configuration property
• Trigger an Event : Add event description
• Proximity Detector : Supports object within object & object contains object proximity types, can output proximity groups as object attributes instead of new objects
• Generic Detector : Improve CPU utilization

Platform

• Search enhancements
  • Bulk select & delete search results
  • Thumbnails for saved snapshots
  • Apply Labels to a file from search result detail view
• Node properties "object label" selection experience improved to support attribute selection as well

Gateway

v1.34.7 (Oct 30)

• Parallelized inference: TensorRT engine generation now handles ONNX with fixed batch (before it was required convert them to dynamic)
• AI Models: added inference parser for Ultralytics YOLOv5 and YOLOv8 models exported to ONNX format
• Generate thumbnails for Snapshot node output images
• Generate thumbnails for cameras/streams snapshots
• ONVIF auth optimizations to reuse digest auth headers and reduce auth frequency when monitoring existing cameras
• Yolo ONNX models now default to maintain_aspect_ratio enabled (letterbox mode)

New Custom Function Repo node enables loading custom functions from Git repositories. Nodes gain UUIDv4 event IDs, attribute modifications, metadata filters for webhooks, and improved object re-identification. Platform updates add file thumbnails, bulk deployment updates, and improved deployment details. Gateway v1.33.29 includes CPU leak fixes, performance enhancements, Triton support for license plates, reduced container sizes, and improved error handling.

Nodes

• 🍀 Custom Function Repo : Load custom functions from a Git repository
• Trigger an Event : Add a UUIDv4 event_id for all Events
• Modify and Merge Objects supports attribute modifications
• Send HTTP Webhook : Add filters for metadata sent in the webhook
• ReIdentify Objects : Prevent duplicate id’s for quick matches

Platform

• Dashboard: Add thumbnails to files widget
• Allow bulk updating deployments from deployments overview
• Display queued deployments info in deployment overview
• Make edit configuration icon into an actual button on deployment detail view

Gateway

v1.33.29

• Fix CPU leak (increasing CPU usage over time)
• Performance enhancements and profiling support
• Retry API requests on "connection error"
• Don't restart file based deployments when stall is detected
• Pre-install tensorflow==2.11 on lumeod containers
• Add triton support to License Plate Recognition model
• Reduce container sizes by 20+%

New Crop and Warp node simplifies video adjustments, while nodes gain support for custom text display, local SDK mode for license plates, and multi-class classification. Platform updates introduce shared stream URLs for external embedding and fix deployment form issues. Gateway v1.31.17 adds Ubuntu 24.04 support, per-node deployment metrics, and resolves deployment state update issues.

Nodes

• 🍀 Crop and Warp node for easier crop / perspective warp of input video
• Display Stream Info : Support to show custom text
• License Plate MMC Platerecognizer : Support local SDK mode
• Generic Classifier: Multi-class classification support

Platform

• 🍀 Shared Streams : Create signed stream URLs that can be shared, accessed and embedded outside Lumeo console
• Fix new (or copy) deployment form not respecting video source selection to limit gateway select options
• Use pipeline values after deploy-time property is reverted to non-overridable

Gateway

v1.31.17

• lumeod installer: Support Ubuntu 24.04
• Introduce per-node deployment metrics
• Fix deployment state update failure after restarting a large number of deployments

New Nvidia NIM Lookup node adds advanced scene and object lookup capabilities. Enhancements include synchronous lookups for license plates, event logging for triggers, fallback labels for classifiers, and attribute probabilities in stream info. Platform updates unify recordings and search views, with enhanced search overlays for specific sources. Gateway 1.30.27 introduces a zero-fps watchdog, container restart support, upgraded Jetson onnxruntime-gpu, and automatic local file cache cleanup when storage limits are exceeded.

Nodes

• 🍀 Nvidia NIM Lookup : Use Nvidia NIMs to lookup scenes and objects using large vision models
• License Plate MMC Platerecognizer :
  • FIx: LPR Platerecognizer add missing LPR region
  • Fix: Minimum lookup size is now respected
  • Enable synchronous lookup for clips
• Trigger an Event node now supports Events logging
• Generic Classifier: Generic classifier supports fallback labels when no classes are found. Useful for negative labels ("no ppe detected")
• Display Stream Info : Show attribute probabilities

Platform

• Unified files + search view to list all recordings and enable search amongst them. Head to Search to use.
• Search enhancement: Files overlay when selecting a search result now filters by that source, making it easy to review events that happened before and after on that specific camera/source.

Gateway

1.30.27

• Zero fps watchdog : Restart deployment if it is stalled for more than X minutes.
• onnxruntime-gpu in Jetson containers upgraded
• Support to restart gateway container from Lumeo console/API.
• Temporary local file cache is deleted when plan cloud storage limits are hit