Case Study: Real-time Physical Activity Detection & Feedback using Wearable Accelerometers, Indoor Positioning, and Smartphones
In early 2011, EveryFit conducted a 2-month physical activity study with employees of a Fortune 500 bank in the Boston area. The objective of the study was to use smartphones, wearable accelerometers (one wrist, one ankle), and indoor positioning to detect physical activity in the gym and provide real-time individual and group feedback on the smartphone.
23 subjects participated in the study and were divided into teams of 3 or 4. The participants were selected from a group that was participating in a formal weight loss competition sponsored by the bank.
Indoor Location Tracking:
To track indoor physical activity of the subjects, EveryFit first needed to pull location data in the gym. To do so, our team fitted the bank’s on-site health club facility with 4 WiFi beacons. EveryFit developed an indoor positioning algorithm that used Sequential Monte Carlo methods and Hidden Markov modeling to identify 4 discrete zones of the gym.
Fitting on-site gym with WiFi beacons
Gym Floor Map: A) Group Fitness Class, B) Cardio Zone
C) Machine Weights Zone, D) Free Weights Zone
With the 4 WiFi beacons installed, EveryFit needed to fit the participants with a device that would communicate with the beacons to sample the signal strength. The device of choice was an Android smartphone, which participants were asked to carry during their workouts. EveryFit installed custom software on the phones that uses EveryFit’s SDK to scan the strength of the signals from each of the 4 WiFi beacons. Through persistent signal scanning, EveryFit was able to infer a participant’s position inside the gym.
A dynamic model of the environment was used to infer a participant’s position inside the gym every 15-20 seconds. Since the time of the study in 2011, EveryFit optimized the algorithm and demonstrated room-level accuracy with as few as 3 access points. A few notes:
1) The algorithm does not assume a particular probability distribution for the measured signal so it can follow a non-Gaussian distribution (e.g. multimodal or multinomial)
2) The estimation of the probability distribution will depend on the number of samples measured but that does not seem to be a barrier since initial finger printing of the different areas took around 2 minutes per zone
3) The gym was more challenging than expected because it was a relatively small gym with a lot of machines.
A demonstration of the positioning technology can be seen in this video:
Lower and Upper Body Activity Measurement:
Next, to detect lower and upper body activity of the subject, EveryFit used wearable accelerometers ("Wockets") developed at MIT. At the beginning of each workout, each of the study participants were asked to pick up a Wrist ("W") and Ankle ("A") sensor and their Android smartphone at a charging station in the gym:
Charging station with wrist and ankle sensors paired to an Android smartphone
Participants were provided with the following instructions for their gym workouts:
Thus, EveryFit was able to stream location data and lower and upper body activity data in real-time to the smartphone. With that data now residing on the phone, EveryFit was able to classify the discrete physical activity type of the participant (e.g. shoulder press, elliptical machine, bicep curls, stair master, etc). In real-time, the phone uses a window-based method to break the data into windows and computes a wide range of features. This data is then fed into a C4.5 algorithm that is used to classify the discrete activity of the user. A video showing the classification is below:
By classifying physical activity type of the participant, EveryFit could develop more accurate objective measurements of energy expenditure, reps, step counts, and other performance measurements.
Participants were able to see two views of their workout progress on the phone:
Study participants would see the size of their trees grow/wilt depending on daily attendance record, activity level, successful completion of challenges, and team performance
Sample real-time intensity challenge
Social Benchmarking View:
For the more graphically inclined participants, EveryFit offered a real-time benchmarking view that provided a measure of energy expenditure vs. their previous best, the member average or the member best (member average/best factored in data from all 23 study participants).
Lastly, participants were able to see a breakdown of their energy expenditure and exercise intensity via a Web Dashboard:
Demonstration of Real-time Physical Activity Detection
The following demo is a real-time simulation from a participant’s workout. The participant was a 54-year old female who starts working out in a group fitness class. At the beginning of her workout, she receives a challenge to do 10 minutes of strength training to earn a 10x bonus for her tree. Once her 1-hour fitness class finishes, you can observe that she changes activities and does slightly more than 10 minutes of shoulder press and leg press exercises.
The moving particles you see in the video reflect the objective intensity level of the participant as well as EveryFit’s confidence in a participant’s location. The faster the particles are moving, the greater the intensity of the participant at a given time. The more concentrated the particles, the greater our confidence that the participant is in a particular zone; the less concentrated, the lower our confidence level.