Jumping Motion Recognition

Research Description

This research was developed at the Figur8. Inc to identify the motion state of people when they jump by detecting data from wearable sensors.

• Purpose. Extract the motion information from wearable sensors, such as jumping up and down.
• Challenge. Simplify finding data with jumping feature quantities in a series of time series data.
• Data. The data comes from a specific muscle sensor, which records muscle stretch during jumping.

Research Method

As shown on the left of the figure, according to the recorded video, mark the video of the jumping part according to the time point. In the modeling process, the LSTM of the RNN is considered as the best model, because the collected data has the characteristics of time series, and the tested LSTM mode is easier to identify the information of motion in the time series. Model information is shown on the right.

• Labeling data. Based on the recorded video marker jumps, label the data of the jump segment as 1 and the rest as 0.
• Build the model and training. The reshaped data is normalized and put into the RNN network for training.

RNN Test Results

Summary

By comparing the predicted results with the marked results, the prediction accuracy is 77%, and the jumping error is within 1.8 seconds.