Get Started
Action Recognition
On this page
  1. Commands
  2. Settings
  3. Output Variables
  4. JSON Output

Action Recognition

Recognizes what a person is doing from 400 action categories (walking, running, sitting, clapping, cooking, and more). Used for fall detection, exercise form checks, and behavior analysis.

Commands

CommandDescription
Specific Action?Checks whether the target action is detected at or above the threshold.
Recognition ResultWrites the name and probability of the most probable action.
All ResultsWrites every recognition result, ordered by probability, as JSON.

Settings

SettingDescription
Target ActionThe action watched by Specific Action?. One of the 400 actions of the Kinetics-400 dataset; the default is abseiling.
Detection Threshold (%)Probability at or above this value counts as the action. 10 to 100, default 50.

Output Variables

FieldTypeValue
Action NameTextEnglish name of the most probable action.
ConfidenceNumberProbability of the recognized action, 0 to 100.
Action MatchedDigitalTrue when Specific Action? finds the target action at or above the threshold.
JSON ResultTextArray of every recognition result.

JSON Output

An array with, per result, label, confidence (0 to 100), and the model class number index, ordered from the highest probability.

[
  { "label": "walking", "confidence": 85, "index": 0 },
  { "label": "running", "confidence": 10, "index": 1 }
]

For the full list of the 400 actions see the official label list of the Kinetics-400 dataset (github.com/cvdfoundation/kinetics-dataset).