Gait metrics help us understand your health. They show your walking speed, step length, and how long both feet are on the ground. These metrics are important signs. They can show your risk of falling. They also show problems with your nerves or muscles. Watching how you walk, called gait analysis, gives helpful health information. It does this without needing to go inside your body.
Before, measuring gait needed costly lab tools. This made it hard to track walking all the time. Now, smartphones can help. They use built-in sensors. But you must place them just right, like in a front pocket or on a belt. Smartwatches are worn on the wrist. They stay in one place. This makes them better for constant tracking. You can even track your walk without your phone.
But smartwatches have not been as good as phones for checking many walking details. Our study, "Smartwatch-Based Walking Metrics Estimation," aimed to fix this. We showed that smartwatches can accurately measure many walking metrics. They work as well as phones. This makes them a good choice for tracking your walk.
A deep learning approach for the wrist
We used a special computer learning method. It is a deep learning model. It works for both smartwatches and smartphones. This model can guess many things at once. It uses a TCN architecture. This is different from other models. Older models only found walking events. You had to guess the distance from that. Our model directly guesses walking details like step length and speed.
Our model needs two main things. It needs your height. It also needs the sensor signals from the watch. This includes data from the accelerometer and gyroscope. We used a Pixel Watch. It sent data 50 times every second. The model learns from these signals. It then adds your height. Finally, it guesses all the walking details. It guesses speed and how long both feet touch the ground. It also guesses step length, swing time, and stance time for each foot. Here are the details:
- Gait speed: How far you walk divided by the time it takes (in cm/s)
- Double support time: The time both feet are on the ground (in %)
- Step length (unilateral): The distance from one foot landing to the other foot landing (in cm)
- Swing time (unilateral): The time your foot is not touching the ground (in ms)
- Stance time (unilateral): The time your foot is touching the ground (in ms)
We used 5-second chunks of data. We overlapped them by 1 second. We used MAPE to check how well the model worked. MAPE measures the average error. It helps the model guess distances and times correctly.