iOS · Android · 设备内计数
先做一遍，
剩下的交给计数。
对着相机完整做一遍动作，设备在本地记住这段轨迹。之后重复该动作时逐次计数，达成时播放提示音并由系统语音朗读次数。
即将上架
iOS · Android · On-device counting
Do it once.
Let the count take over.
Perform the movement once in front of the camera, and the device learns that trajectory locally. Repeat it, and every rep is counted — each one marked by a tone and the number spoken aloud.
Coming soon
运作方式
三步，全部在
这台设备内完成。
姿态识别在设备内完成；iOS 订阅购买与验证可能需要联网。
-
示范一轮动作
对着相机完整做一遍，让参与动作的部位持续入镜。单次示范最长 30 秒。
-
确认起止范围
以缩略图胶片条裁剪出「起点—动作端点—回到起点」的完整周期。
-
开始计数
重复该动作时按轨迹顺序完成一次匹配才计数，达成时播放提示音并由系统语音朗读次数。
How it works
Three steps, all of
them on the device.
Pose recognition runs on the device; iOS subscription purchases and verification may need a network connection.
-
Demonstrate one rep
Perform the movement once, all the way through, in front of the camera, keeping the parts of the body it involves in frame. A demonstration can run up to 30 seconds.
-
Confirm the range
Use the thumbnail filmstrip to trim one complete cycle: the start position, the far end of the movement, and the return to the start.
-
Start counting
As you repeat the movement, a rep counts only once it has matched the trajectory in order. Each count plays a tone, and the number is read aloud by the system voice.
功能
从示范，
到报数。
-
设备内学习
从单次示范中提取动作周期，重采样为固定长度模板，全部在设备内完成。
-
骨架叠加显示
在相机预览上绘制识别到的关节与骨段。
-
动作保存与复用
已学习的动作以 JSON 保存在应用文档目录，可在首页选择既有动作直接开始计数。
-
语音报数
计数达成时播放提示音并由系统语音朗读次数。
-
前后摄像头切换
倒计时与录制阶段禁止切换，其余阶段可用。
-
无 ActiCount 账号，无广告
不提供应用账号注册与登录，不含广告、行为分析或云端同步组件。iOS 订阅通过 Apple 购买与验证。
Features
From the demo
to the spoken count.
-
On-device learning
The movement cycle is extracted from a single demonstration and resampled into a fixed-length template, entirely on the device.
-
Skeleton overlay
Recognized joints and limb segments are drawn over the camera preview.
-
Saved movements
A learned movement is stored as JSON in the app's documents directory, and an existing one can be picked on the home screen to start counting straight away.
-
Spoken counts
When a rep is counted, a tone plays and the number is read aloud by the system voice.
-
Front and rear camera
Switching is blocked during the countdown and the recording, and available at every other stage.
-
No ActiCount account, no ads
There is no ActiCount sign-up or sign-in, and no advertising, behavioral analytics, or cloud-sync component. iOS subscriptions use Apple for purchasing and verification.
当前模型基于人体姿态关节点，适合深蹲、俯卧撑一类全身大关节动作，不适合依赖手指动作、器械操作或物体交互细节的动作。
The model works from human pose joints. It suits whole-body movements around large joints, such as squats and push-ups, and does not suit movements that depend on finger detail, equipment handling, or interaction with objects.
隐私
画面不出这台设备。
- 相机画面仅在设备内实时处理，不录制视频、不写盘、不上传。
- 示范过程中的 JPEG 缩略图仅存在于当前会话内存中，最多保留最近 12 秒、128 张，用于裁剪确认，会话结束即释放。
- 持久化内容仅有已保存的动作模型（姿态轨迹数值，不含图像）、语言偏好与录制显示偏好。
- 姿态推理由打包在应用内的 RTMPose 权重完成，无后端服务。
- Android release 构建仅保留相机权限，网络等权限已在清单中移除；iOS 仅声明相机用途。
Privacy
The video never
leaves the device.
- Camera frames are processed on the device in real time — no video is recorded, written to disk, or uploaded.
- The JPEG thumbnails captured during a demonstration are held only in the memory of the current session — at most the last 12 seconds, up to 128 of them. They serve to confirm the trim, and are released when the session ends.
- The only things kept on disk are saved movement models — numeric pose trajectories, with no images — plus language and recording display preferences.
- Pose inference runs on RTMPose weights bundled inside the app. There is no backend service.
- The Android release build keeps only the camera permission, with network and the rest removed from the manifest; on iOS, camera use is the sole declaration.
取得应用
即将上架 App Store。
系统要求：iOS 16.0 或更高版本；Android 需要相机权限。界面提供 12 种语言，含简体中文与英文。
Getting the app
Coming soon to
the App Store.
Requires iOS 16.0 or later; on Android, the camera permission is needed. The interface is available in 12 languages, English and Simplified Chinese among them.