Skip to main content
Version: 2.8.1

Human action recognition (HAR)

HAR (Human Action Recognition) includes recognition of human falls, fights, lying and sitting in video streams and video files.

System requirements​

The following configuration is required to enable HAR on a single Full HD @ 25fps video stream:

Minimum system requirements +

  • GPU: performance ~FP32 10 TFLOPS; 3GB VRAM; CUDA11 Runtime;
  • RAM: +4Gb;
  • CPU: 2 AVX cores @3GHz.

When an action type is detected, OMNI Agent generates a process and an event which can be sent to OMNI Platform or to a third-party service.

How to enable HAR​

Open the run_params.json configuration file and set "true" in the fields below:

  • events.har_events.fall.enable — fall
  • events.har_events.fight.enable — fight
  • events.har_events.sit.enable — sit
  • events.har_events.lie.enable — lie
info

When enabling HAR, we recommend using:

  • CUDA: set "true" in the use_cuda_onnx field of the run_params.json file or enable "CUDA usage" option in Settings of OMNI Agent web-interface.
caution

For stable work of OMNI Agent, disable the generation of face biometric templates. To do this, open the run_params.json configuration file and set "false" in the pipelines.face.enable_templates_generation field.

img.png