accelerometers have been used for a variety of applications; we're interested in seeing what they can tell us about dancers

We can collect data from a variety of sources (dance teams on campus, experienced dancers vs. non-dancers) and see if we can find distinct patterns. First, however, we have some questions to investigate, like:

• What is a dance? Can we identify dancing as a specific activity?
• Can we distinguish these dance styles from each other?
• How do we represent accelerometer data to other humans in a way that's useful and meaningful?
• How do you do this in real time? Dancing happens over time, so a single data point doesn't tell us much about the overall activity.

We started our project by just collecting lots of different accelerometer data. this set us up for doing exploratory data analysis (EDA). Very early on, we realized that our project could go in a variety of directions; specifically, our project had components of classification (e.g., walking vs dancing), visualization (e.g., how do we represent accelerometer data for humans), and familiarization (i.e., learning how to work with accelerometer data in the first place).

We primarily used an app called SensorLog to collect data. We faced a couple of hiccups with this, since SensorLog seems to have some errors in data collection that we can't really explain. We managed to get access to a home-grown accelerometer app, but faced the issue of our data having different formats to work in.

Figure: example of oddities in SensorLog data. The x-axis is the sample, and the y-axis is the associated time. We would expect a smooth line (as in the gyroTimeStamp) but sometimes see the strange steps for logging time. This is the same data source!

Overall, however, our project started to look promising fairly early. After some less-than-promising visualizations like the spectrogram below, we were feeling concern; some minor signal processing helped make our data much more promising, though.

As our data cleaning became more sophisticated, we also began to collect more intentional data. In order to deal with the problem of classification, we collected data for line dances (the wobble and the cupid shuffle(, structured dance (zumba and salsa), and motion data (walking and running). This data was usable for our work with visualization.

Some of our key findings were (see poster at end for analysis details):

We used a low pass filter to help visualize and analyze the accelerometer data, shown below. The time domain graph shows us which pocket the accelerometer is in.

We then explored the data in the frequency domain, below.

We can see that an activity like walking provides much more of a regular cycle than a dance like salsa. These data also lead us to...

This graph shows distinct clusters between walking activity and dancing data. This is particular data set doesn't distinguish dances from each other, however, and we need to collect a larger number of samples and refine our feature extraction to make this meaningful.

Accelerometer data can be displayed in three dimensional space to provide a computational representation of where the accelerometer is in space. So, we have some sense of what a dance physically looks like, which is a step in the direction of visualizing this information in a human-consumable manner. The visualization above is a sped-up version of a person (note how we can see the person's altitude change as they walk uphill.)

raw data from SensorLog

The node server allows us to pipe SensorLog data over a local network; this sets us up for future work in real time visualization of sensor data. In particular, this cleans the data up in a easily readable JSON format - this could be stored in a database, or accessed for immediate processing and visualization.

Overall, some of our basic questions seem to be answered:

• What is a dance? Can we identify dancing as a specific activity? Yes! Dance is a specific type of motion, and accelerometer data can be used to differentiate from other common behaviors that occur alongside dance - like walking.

• Can we distinguish these dance styles from each other? For the moment, no. Dance could probably be described as "rhythmically frenetic." So, while we can distinguish dancing from walking, there's simply too much noise to make a meaningful differentiation between the different types of dance with our current data set.

• How do we represent accelerometer data to other humans in a way that's useful and meaningful? We have a lot of work to do on this front. Our visualizations have primarily worked as a proof of concept, and our current visualization seems intimately tied to the realistic motion, but this doesn't necessarily feel like it adheres to a meaningful principle of information visualization - we'd have to consider something like pre-attentive processing to see whether it actually means something to people.

• How do you do this in real time? Dancing happens over time, so a single data point doesn't tell us much about the overall activity. This is a great question, past-selves. We'll get back to you at the end of P2. (:

Moving forward, we need to investigate that real-time work, but we could do a lot to retest the work we've done more rigorously. There's a lot of opportunity to collect data, and the representational work is still in its infancy. We also might want to start considering the ability to fool the sensors - and, realistically, thinking about this in terms of its applications. This concept is exciting, but also easily misapplied. We have to do a lot of consideration for how this would work in real-world situations. That being said... We could also do a lot to even start cleaning up our code and files, but that's less exciting, so we probably won't get around to it just yet.

thanks for reading! we've also included a poster below