Skip to main content

Detection Scenarios

The first step in starting AI analysis in EVA is to create a Detection Scenario. A detection scenario goes beyond simply identifying objects—it defines rules that allow the AI to interpret situations and trigger alerts only under desired conditions.



Refining Scenarios​

The more specific your detection scenario is, the more accurately the EVA Agent can understand the situation. In particular, to reduce false alarms, it is important to clearly define Excluded Situations where alerts should not be triggered.

By continuously reviewing triggered alerts and refining your scenarios, EVA becomes progressively more accurate and delivers highly reliable results with minimal unnecessary alerts.



Creating Scenarios with the Generate Feature​

EVA provides a Generate feature to assist users in writing detection scenarios.

During this process, the internally operating Enrich Agent automatically designs scenarios based on user input.

Overview of the Enrich Agent​

The Enrich Agent acts as a “designer” that transforms abstract user requests into precise detection scenarios that AI can understand. It analyzes the input text and determines whether simple object detection is sufficient or if deeper contextual understanding is required, then generates an appropriate structure automatically.

Core Logic​

  • Inference Type Determination

    • VM Only: Scenarios that check whether an object appears in the frame
    • VM + VLM: Scenarios that require multi-step, context-aware reasoning

  • Scenario Optimization

    • Analyzes the input and clusters/splits it into up to 4 scenarios

  • Scenario Structuring

    • Goal & Case Generation

      • Defines detection objectives and extracts key situations used for alerts and feedback filtering

    • Frame Mode Selection

      • Single: Based on a single image
      • Sequence: Based on 3 consecutive frames where the object is detected (within 5 seconds)
      • Timeline ($T$ sec): For time-based scenarios, uses 3 evenly spaced frames within $T$ seconds

    • Detection Steps

      • Step-by-step conditions (alerts are triggered only when all conditions are True)

    • Excluded Situations

      • Conditions where alerts must be suppressed (alerts are triggered only when all exclusion conditions are False)

Generation Workflow​

  1. Input Draft: Enter a brief text description of the desired situation.

  2. Click Generate: The EVA Agent analyzes the input and automatically creates a scenario including Detection Steps and Excluded Situations.

  3. Refine and Improve: Enhancing the generated scenario leads to significantly more precise detection results.

💡 TIP: Scenarios generated by EVA are drafts. Continuously refining them based on actual alert results allows optimization for real-world environments.



Detection Scenario Generation Examples​

User Query: "Notify me when an ambulance appears"

Inference Type: VM

Case: Ambulance detected

Frame Mode: Single

Detection Steps:

  • Step 1: An ambulance is visible in the frame.

Excluded Situations:

  • Cases where it is not possible to confirm the presence of an ambulance
  • When typical ambulance features (yellow/white body + emergency lights) are not clearly visible, or visibility is impaired due to blur, darkness, or obstruction
  • When all yellow/white vehicles in the frame are recognized as general vehicles or other types of emergency vehicles (e.g., fire trucks)


User Query: "Notify me if a person is lying on the ground"

Inference Type: VM + VLM

Case: Person collapsed

Frame Mode: Single | Sequence

Detection Steps:

  • Step 1: A person is observed in the frame.
  • Step 2: At least one person is in a collapsed state.

Excluded Situations:

  • All individuals who appear collapsed are actually resting or lying on chairs/sofas
  • The leg area of all individuals is partially out of frame, making it difficult to determine whether they are collapsed
  • The leg area of all individuals is significantly occluded by other objects, making it difficult to determine whether they are collapsed


User Query: "Notify me if a person is loitering for more than 60 seconds"

Inference Type: VM + VLM

Case: Person loitering for over 60 seconds

Frame Mode: Timeline (60 sec)

Detection Steps:

  • Step 1: At least one person is observed in the frame.
  • Step 2: The person is moving freely outside a defined boundary or work area.
  • Step 3: The person is exhibiting loitering behavior.

Excluded Situations:

  • All individuals who appear to be loitering are actually stationary or staying in a fixed position
  • A significant portion of the person’s body is out of frame, making it difficult to determine loitering behavior
  • A significant portion of the person’s body is occluded by other objects, making it difficult to determine loitering behavior


Automatic Extraction of Detection Targets​

EVA automatically extracts detection targets from user-written sentences, defining what the model should detect. This helps clarify detection objectives and improve model recognition performance.

Example Input: "Notify me if there is a fire or even a small amount of smoke"

Detection Targets: [Fire, Smoke]



Situation Analysis and Alert Decision​

When specific conditions are detected based on the defined scenario, the EVA Agent collects image snapshots according to the frame configuration defined in the scenario (single or multi-frame).

It then uses a VLM to analyze the situation by incorporating temporal flow and contextual understanding.

Based on this analysis, EVA determines whether to trigger an alert, and users receive alerts with detailed messages.

If unnecessary alerts occur, adjust the Detection Steps or Excluded Situations.

Through continuous refinement, EVA becomes increasingly precise and delivers highly reliable decisions.