History of Activity Tracking
First, let’s define activity tracking:
ACTIVITY TRACKING PRIMARILY REFERS TO RECOGNIZING USER ACTIVITY/MOVEMENTS AND RECORDS A PERSON’S FITNESS ACTIVITY
One of the first activities that were attempted to recognize was walking.
- All the way back in the 15th century Leonardo da Vinci, sought to track the distance a Roman soldier walked. He drew plans to develop a device and the first pedometer was recorded.
- In 1780, the next piece of pedometer was made by Abraham-Louis Perrelet, a credit Swiss watch maker, who made watches that were motion sensitive and self-winding, with inventing the mechanism that would be used in later pedometers to record steps.
- In 1965, a man named Y. Hatano from Japan put forth a pedometer called the “manpo-kei” which in English translates to 10,000 steps meter.
- By 1985 research by Hatano himself was widely accepted as to showing that 10,000 steps a day was the proper energy output for a person to balance their caloric intake to maintain a healthy weight. In Tokyo, these pedometers were recognized to be accurate, and were named “manpo-meter”. They claimed to be the world’s first device to measure the number of steps of walking.
More sophisticated pedometers, work entirely electronically and, since they have no moving parts, tend to be longer-lasting, more reliable, and considerably more accurate. They dispense with the swinging pendulum-hammer and measure your steps with accelerometers instead. These are microchips arranged at right angles that detect minute changes in force as you move your legs. A computer evaluates whether the movements captured resemble a step and a calculation is made. Activity trackers are fundamentally upgraded versions of pedometers, in addition to counting steps, they use accelerometers and altimeters to calculate mileage, graph overall physical activity, calculate calorie expenditure, and in some cases also monitor and graph heart rate and quality of sleep.
Now-a-days activity trackers act like a personal coach, where it automatically recognizes physical activity and calculates calorie expenditure. It also keeps a track of it and generates exercise statistics and infographics. They ensure that you are able to focus on enjoying better health and getting in shape through tracking the steps you take, calories burned and your heart rate. Though activity trackers help people in staying fit, they couldn’t provide longer battery life, which is its biggest drawback. Due to this, one has to monitor battery life of activity tracker continuously. Sometimes people may even miss activity tracking, due to the shorter battery life of activity tracker. We came up with a solution which provides one year battery life and more accurate tracking that helps the user avoid monitoring battery life of the activity tracker.
Automatic Activity tracking using MUSE smart watch
For automatic activity recognition, we used supervised machine learning algorithm. In supervised machine learning, the machine is trained with labelled data and an optimized model is built from the data in predicting the activity. Hence, the accuracy of the model majorly depends on the data we train. Apart from the data, feature extraction, feature selection from data and algorithm selection for training the machine add value in prediction of activity.
For better quality of data, We distributed our smart watch to athletes and sports persons and asked them to perform certain tasks while wearing it. Next, we collected feedback and data from them. This made sure that we gained a large amount of data to train the machine, ensuring accuracy.
Unless we have enough data, the machine may not be capable enough to build the complex model which is required to predict all the activities. It leads to ‘Underfitting’. In this case, the machine shows a high bias but low variance. However, if you train the model with too much data, the machine captures noise of the data and tends to fit all the data points. This leads to ‘Overfitting’. The machine may perform well on the trained data but performs poorly on new data sets. In this case shows high variance but low bias.
In order to avoid underfitting and overfitting, we have performed cross validation. It is a technique, which involves reserving a particular sample of data, on which you do not train the model. After building the model we use this sample of data as a test set, which helps in finalizing the model.
It is always difficult for the machine to recognize the activity from raw accelerometer data. So, we have performed feature extraction and extracted more features from the raw accelerometer data, for the machine to understand and recognize activities with more ease and accuracy.. It is a technique which involves deriving new variables from the existing variables to unleash the hidden relationship of data set.
To avoid the impact of one feature on the whole model, we have performed ‘Feature Scaling’. All features are normalized using this technique. It also helps in achieving the optimized model faster.
In order to make the model more efficient, we have performed feature selection. It is a process of finding out best set of independent variables in predicting the dependent variable. Here we selected best set of features by visualizing relationship between variables and by performing statistical tests.
After performing above steps and by choosing the right machine learning algorithm, we obtained results accurately. Selection of algorithm mainly depends on nature of problem, number of features and size of the data.
Why Muse Wearable?
Although there are many fitness trackers in the market which help in tracking activity but they need to be charged once in every two days. So, people have to always monitor their battery level of fitness tracker. Due to this some people miss tracking their activities continuously.
Muse Wearables is providing a smart analog watch, which has a battery life that lasts for one year and is more accurate in activity tracking, all at an affordable price!