Getting Started

Core Function Overview

The sensor supports 3 recognition modes + one‑click learning function, as described below. This section briefly introduces the functions and precautions for each mode.

1. Custom Gesture Recognition

Recognizes gestures based on 21 human hand keypoints. Users can learn and save up to 8 custom gestures. Data is stored in the on‑board Flash and is not lost when power is off. The gesture recognition similarity threshold can be adjusted freely. Gestures can be deleted or re‑learned at any time to meet personalized interaction needs.

Precautions:

• Recommended single‑gesture learning time: 3~5 seconds. Keep the gesture clear and unobstructed during learning.
• It is recommended to set the recognition similarity threshold between 70%‑80%. Too low a threshold may cause false recognition; too high may make triggering difficult.
• Only one hand gesture can be recognized and learned at a time.
• For complex gestures, it is advisable to learn multiple times and keep the best result. Recognition accuracy can be improved by adjusting the similarity threshold.
• The sensor can output real‑time coordinates of the 21 keypoints. You can also calculate joint angles based on the coordinates to implement more complex gesture judgments.

2. Custom Posture Recognition

Based on 17 human body keypoints (head, shoulders, elbows, wrists, hips, knees, ankles, etc.), supports static posture recognition for a single person. Users can learn custom fun postures, such as stretching, standing on one leg, raising a hand, etc. Suitable for interactive games, non‑contact control, smart interaction, and similar applications.

Precautions:

• Optimal recognition distance: 2 meters. Try to keep the entire body clearly visible in the camera frame; the more keypoint data, the more accurate the recognition.
• Only one person’s posture can be learned at a time, but the sensor can recognize skeleton data of multiple people.
• For complex postures, it is advisable to learn multiple times and keep the best result. Recognition accuracy can be improved by adjusting the similarity threshold.
• The sensor can output real‑time coordinates of the 17 keypoints. You can also calculate joint angles based on the coordinates to implement more complex posture judgments.

3. Fixed Gesture Recognition

The sensor has 13 pre‑programmed fixed gestures that require no prior training. They can be recognised immediately after power‑on, making them suitable for quickly building basic gesture‑interaction projects. Supports common gestures such as like, dislike, OK, numbers 1‑5, stop, etc.

Precautions:

• This feature is still in the testing phase and recognition accuracy is currently limited. We are continuously optimising the AI model.
• This feature has high requirements for background environment. Recognition works better on a clean white background, but performs poorly in complex background environments.

4. Learning Function

The sensor’s learning function is the core underlying capability for realising custom gesture recognition and custom posture recognition. All training processes are completed locally on the sensor, without requiring an internet connection or cloud computing. Beginners can get started quickly.

Precautions:

• Environmental requirements: Learning should be carried out in a well‑lit, evenly lit environment. Avoid backlighting, direct strong light, or dim conditions. Keep the background as simple as possible to minimise interference from irrelevant objects or people.
• Action requirements: Keep the action stable and standard during learning. Do not shake or make unnecessary small movements. Ensure the sensor can fully capture the core features of the gesture/posture.
• Differentiation principle: The differences between different custom gestures/postures must be sufficiently obvious to avoid cross‑misrecognition due to similar features.
• Storage limitations: A maximum of 8 custom gesture models can be saved at the same time. The number of custom posture models is limited by the on‑board Flash capacity. It is recommended to save only as many as needed.
• Status indication: The sensor provides a learning failure prompt during the learning process; follow the prompts accordingly.

Host Computer Software Tutorial

This tutorial aims to help users, after understanding the sensor’s functions, quickly experience the sensor’s core capabilities using the companion host computer software.

1. Hardware Preparation

• Gravity: AI Posture and Gesture Sensor (SKU: SEN0670) ×1
• USB Type‑C cable ×1
• PC ×1

2. Software Preparation

• Download SEN0670 host computer software: SEN0670_Upper Computer Software.zip
• Download CH343 Driver: CH343SER Driver.zip

3. Hardware Connection

4. Operation Steps

1. Download the SEN0670 host computer software, save it to a local folder, and unzip it. Locate the .exe file as shown in the figure below.
   

2. Double‑click the .exe file to run the host software. The host software interface will appear as shown below.
   

3. After connecting the sensor device, the host software will automatically detect the device port number. Simply click the Open button. The camera view and data from the sensor will be displayed, as shown below.
   

Core functions of the host software:

4. The host software defaults to gesture recognition mode. As shown below, make any gesture. The sensor will recognise it, output the corresponding posture ID as 0, confidence 93, and output the real‑time data of the 21 keypoints of that gesture.
   

5. Next, select posture recognition mode. The sensor does not need to be powered off; the change takes effect immediately. As shown below, make any posture. The sensor will recognise it, output the corresponding posture ID as 0, confidence 90, and output the real‑time data of the 17 keypoints of that posture.
   

6. Finally, try the fixed gesture recognition mode. As shown below, make any gesture according to the gesture types shown on the right. The sensor will recognise the corresponding gesture ID and name.
   

Note: This mode is still in the testing phase and recognition accuracy is currently limited. It performs poorly especially in complex backgrounds – for example, it works better on a clean white background.

7. Learning operation: Take gesture recognition as an example. The operation for posture recognition is the same. Switch to gesture recognition mode.

   When a custom gesture is detected in the camera frame, click the Learn button. Keep the gesture still for 3~4 seconds until learning is complete. The colour of the gesture will then change.
   

   After learning is complete, the learned gesture appears in the learned content area. The default ID is 1 (IDs accumulate sequentially; maximum 8 gestures can be learned, so the maximum ID is 8). The default name is Gesture_1. The ID name can be renamed (only character naming is supported). After saving, the change takes effect.

   For example, rename it to like. After saving, make the same gesture again, as shown below. The recognition result shows ID:1 like with a similarity of 88% to the defined gesture. The gesture has been successfully learned.
   

   Repeat the above steps to teach the sensor more custom gestures or postures. The ID numbers or ID names of learned gestures/postures can be obtained on the host side and used to write corresponding execution programs.

   Notes on using the host software:

   1. Use the Clear button with caution. Pressing it will erase all learned gestures or postures.
   2. Optimal distance for gesture recognition: 50~60 cm. Minimum distance for posture recognition: 2 m.

8. Other functions of the host software: Click the Settings button.
   
   The following functions are supported:

   1. Firmware upgrade: Update the firmware
   2. UART baud rate: Set the communication baud rate for the UART interface
   3. About: View product related information