Skip to content

Release notes - CVEDIA-RT 2022.11.1

Solutions

Security

The Security solution has new capabilities.

Person Vehicle Animal (PVA) classifier

It's now possible to classify motion detected objects using a Person Vehicle Animal (PVA) classifier. This allows skipping the execution of the PVA detector, which could be resource-intensive on some embedded devices.

Classification Options

Movement-only events

Unknown objects moving across a line or entering/exiting a designated area can now trigger events.

Info

Enable the "Movement detection" option on the zone or line that should respond to unknwon moving objects, and disalbe detection and classification in the config panel of the solution.

3D Bounding Boxes

Additionally, you can now enable the 3D Bounding box classifer. This exposes an additional anchor point option to lines and areas for determining with more precision the position of a vehicle inside an area, for example a parking space, even at lower camera angles.

3D Bounding Boxes

Info

You can test this new feature by running the "parking_lot" instance that shows a timelapse of a parking lot. You can also enable the export of zone statistics that reports the number of empty and occupied parking spaces.

Engine Updates

Improvements

  • Allow turning on/off detections on objects
  • Line crossing and areas triggering on any unkown motion detected objects
  • Video codec automatic resolve
  • Shorten the fallen person videos
  • Events dont contain timestamps relative to the source
  • Allowing motion detection even when no trigger is set
  • Figure out Orin deployment requirements / GUI run strategy
  • Port SigmStar plugin to the new code base
  • Ensure softmax is the last layer of classifier models
  • Include in the release headers and sample code for custom plugins
  • Introduce a sleep of 1 millisecond in all solutions inside onRun in case of missing frame data
  • gstreamer local plugin auto discovery doesn't mount local files
  • Update OpenCV dependencies without CUDA in SDK redist
  • Detection settings should be global and not per trigger
  • GStreamer should automatic build a working pipeline regardless of the decoder needed
  • Add h265 / mkv and others gstreamer plugins on Windows
  • Added support for custom model loading (YoloV⅚/7/X)

Bug fixes

  • Security: Instances don't respect detector regions
  • Security: object_area_intrusion doesn't work on linux
  • Security: Adding Object Left Area halt the instance
  • Security: Motion seems in use even when the option is unchecked, and it triggers a reset of the detector options
  • Tailgating Line Event causes crash
  • Benchmark CLI: Hailo segfault
  • MotionPlugin lua signatures are not correct
  • runinstance returns error code even when it runs successfully
  • NetImGUI Fails When running instance
  • JSON + Crops with events - Filenames not consistent
  • Default size of some panels is very small
  • ONVIF: Fast clicking on discover cameras causes segfault
  • Selected detections highlights multiple detections.
  • Changing the settings of one tripwire affects the other
  • Custom sinks used to export data get overwritten by the export settings panel when it's open
  • TextEditor crashes when .json can not be tokenized (happens with complex URLs)
  • Invalid characters in JSON crash RT when opening the Edit Instance window
  • Stopping an instance with a broken gstreamer configuration crashes
  • Base aarch64 image doesn't contain gstreamer
  • OpenCV Dependencies shouldn't be compiled with CUDA