Custom Detection Model

Video analysis object detection provides YOLO series, VisDrone, Road Damage, Smoke Fire, and fifteen other models.

It also supports using custom detection models, including externally trained models and self-trained models. To generate custom models, first obtain sample libraries through the Manage Picture Sample feature, then use the Model Training tool to train sample libraries.

Below are detailed explanations of related content:

Steps

Prepare Sample Library

  1. Create Sample: Click Start tab -> Browse group -> Sample drop-down button -> Picture Sample management button. In the Create/Open Sample Library dialog:
    • Sample Library Name: Enter sample library name
    • Create Sample Type: Select from Image Classification or Image Label (choose Image Label here)
    • Original Picture Path: Select folder containing sample images (supports JPG, PNG, TIFF files in root/subfolders)
    • Minimum Size: Set width/height thresholds (pixels) to filter undersized images

      After configuration, click OK to create sample library. The left panel will display sample list with first image shown in preview.

  2. Image Label: Label detection samples in Sample Picture Window using AI Label, polygon, line, or batch drawing tools. For details, see Image Label Sample.
  3. Export Sample Library: Click Manage Picture Sample tab -> Sample group -> Export Sample Library:
    • Sample Library Name: Use English characters only
    • File Path: Set save location
    • Export Type: Choose Carry Data for large datasets
    • Purpose: Select Video Object Detection
    • Image Scaling Factor: Keep default unless resizing needed
Notes:

Typically requires thousands of sample images for higher model accuracy.

Model Training

  1. After exporting sample library, use Toolbox -> Machine Learning -> Video Analysis -> Model Training tool.
  2. Select samples folder and configure parameters in dialog. See Model Training Parameters for details.

Add Custom Detection Model

  1. After training, add model for video detection: Video Analysis tab -> Transport Analysis group -> Detection Settings.
  2. In Model Settings dropdown, choose Custom -> Add:
    • Model Algorithm: Match training algorithm
    • Model File: Select best.pt from weights folder
    • Type File: Choose *.txt file with same name as sample library
    • Training Image Size: Keep consistent with training setting (multiple of 32)
    • Model Name: Auto-filled or customizable
    • Model Description: Auto-filled or customizable
  3. Alternatively, use Import button to load .sdm model files directly.

This completes the workflow from sample preparation to custom model deployment.

Notes:

Product package path must not contain Chinese characters.

Related Topics

Environment Configuration

Object Detection

Target Statistics

Target Tracking

Geo-fencing Analysis

Velocity Analysis