Body action counting feedback mechanism SDK
This SDK uses AI to detect improper form in 15 types of exercises and provides real-time feedback.
Starting May 12, 2026, Vision Intelligence Platform will no longer issue new licenses for its offline SDK products. Existing users can continue to use the services they have purchased. No new licenses will be granted after this date.
Advantages
High flexibility: You can output information for each of the 15 human keypoints individually, or bind multiple keypoints for fitness action counting. The average counting accuracy exceeds 95%.
Strong adaptability: Template-based configuration allows you to quickly adapt to new movements. The SDK supports real-time negative feedback to help users instantly correct their form. It delivers stable performance across various scenarios, lighting conditions, and backgrounds.
High efficiency: The algorithm is easy to use and integrates quickly. It ensures high frame-to-frame stability with minimal CPU and memory consumption, achieving over 20 FPS on entry-level devices.
Use cases
Fitness apps: Use the device camera to collect data and count exercise repetitions in real time. It can serve as a training assistant, allowing users to review their form and accuracy during playback.
Fitness mirrors: Capture user movements through the mirror's camera for real-time repetition counting. Users can make immediate adjustments based on negative feedback and their own reflection, improving workout quality.
Prerequisites
You must activate and authorize the SDK before use. For more information, see Activate and authorize offline SDKs.
Billing
For pricing details for the Body Action Counting Feedback Mechanism SDK, see Offline SDK pricing.
Supported platforms
Android, iOS
Limitations
Android: Requires Android 5.0 or later. The device must have a camera and a gyroscope module.
iOS: Requires iOS 10 or later. Supported on iPhone 5s and later models. The device must have a camera module.
Recommended camera resolution: The native camera resolution should be greater than 640×480 pixels and less than 1280×720 pixels. Higher resolutions may reduce SDK performance.
If you need to use the SDK on models earlier than iPhone 5s, or on low-end Android devices, please contact us.
The SDK supports only one person in the frame. Their full body must be visible and occupy more than 1/4 of the screen.
Input video resolution: Between 640x480 and 1280x720 pixels.
On mobile devices, the user must face the camera.
Integration
This SDK does not transmit any data to the SDK provider. Therefore, if you use this SDK to collect personal information, you are considered the data processor. The SDK provider is not a data processor or a sub-processor. You must fully inform users and obtain their authorization in accordance with relevant laws and regulations. When informing users, ensure that the SDK provider is not identified as a data processor or a sub-processor.
Android: Android integration guide.
iOS: iOS integration guide.
Appendix
Human keypoint diagram
Supported exercises
Portrait mode exercises: Jumping rope, squats, jumping jacks, squat jacks, star jumps, straight-arm jumping jacks, rhomboid squeezes, arm circles, and torso twists.
Landscape mode exercises: Sit-ups, crunches, push-ups, knee push-ups, and glute bridges.
Feedback message list
Open-source models
Explore and download free open-source models on the ModelScope community: HRNet-2D Keypoint Detection and Whole-body Keypoint Detection - General-purpose Field - 2D.