Pulse Jacket

led_image_2 01This was one of my first Arduino projects. After some near misses with bicyclists while running at night, I decided to get some lights so people could see me in the dark. But why stop at boring plain lights? Wouldn’t it be cool if they could respond to my heart rate?

I looked at a number of existing heart rate sensors for Arduino, but most were optical and could not get accurate readings while I was running since they were constantly being jarred. Since I run with a Garmin GPS watch and heart rate monitor, I tried to hack into the information being sent between the heart rate monitor and the Garmin watch.

Reading a bit more about the technology, I learned that Garmin used the ANT protocol for communication between the watch and heart rate band. The good news was that SparkFun made an ANT transceiver breakout board. The bad news was that the board was discontinued and I could only get my hands on one board. I decided to move forward with this board for prototyping knowing that I would need to come up with a different solution when I made the final project.

The first step was to get the Garmin heart rate monitor and an Arduino communicating with each other. The ANT protocol documents are pretty thorough and they make great bedtime reading. Fortunately for those of us who are impatient, this thread on the SparkFun forums has sample code that already implements the protocol for the Garmin heart rate monitor.

Microcontroller with ANT breakout boardNow that I had the pulse rate information, it was time to add lights. I am a huge fan of Adafruit’s LED strips. These strips have weatherproofing so it would be possible to run outside in the rain. I trimmed the strips to the length of my arms and sealed the ends.

Microcontroller with lightsI added seven different light modes which increased in speed with the heart rate: rainbow, raindrop, range pulse, color shot, twinkle, circulatory and Cylon. Most of these modes are self-explanatory. The range pulse mode faded the strips in time with the pulse and also chose the color based on the current pulse rate (blue being low pulse, red being high). Here you can see a quick demo of the seven modes:

I then began building the final version. For this, I chose to use a Teensy 2.0 because of its low price and small size. I also had to revisit the ANT transceiver. Searching around, I found this ANTAP281M5IB  module with an on-board antenna. After some very delicate wiring and soldering, this proved to be a direct replacement for the SparkFun board.

IMG_8863 Once everything was working, it was time to put this into a portable package for running. The main concern was power. After a bit of research, I found these Energizer power packs that I could plug directly into the Teensy. The one amp power pack would power both LED strips for about an hour. After verifying that everything still worked, I placed the assembled project into a small project box.

IMG_8862The last issue was how to attach the LED strips to my arms. I thought about embedding them into a jacket by sewing them in, but I decided against that as it would be a pain to clean. Finally, I just glued some cable clips to the back of the LED strips and used velcro straps to adjust them for the right fit.

Assembling the arm supportsAssembled JacketAnd after all that we were ready to go! My first real-world test was the Midnight Run in Central Park on New Years Eve 2013.

New Years Eve - 2013New Years Eve - 2013And now I can easily be seen in the dark!

Probabilistic Postal Address Elementalization

I haven’t posted in a few months because most of my time has been consumed with work and school. With respect to school, I have been taking Statistical Natural Language Processing at New York University. As my final project for this class, I have been working on something which I have been curious about for quite some time – probabilistic postal address elementalization.

What exactly does “postal address elementalization” mean? This is the process of breaking a postal address into tokens and classifying the function of each token, such as house number, street suffix or zip code. For my experiment, I created twelve distinct classes: five for administrative areas (country, state, county, city and neighborhood), one for postal codes and six for street address components (house number, prefix, pre-directional, base street name, post-directional and suffix). An example of an elementalized address is below:

address_elementalizationAlthough this seems like a straightforward problem, it is complicated by the fact that many countries have different languages and address formats. For example, in Ghana and Cameroon, there is no standard postal code system. In the Netherlands and Ireland, there is no province or state in the address. In some countries, such as France and Mexico, the postal code is placed before the city whereas other countries place it after the city. Furthermore, known names of terms in specific classes can overlap, so disambiguating streets can be tricky, like Avenue N in Brooklyn (“N” could be a post-directional or street name) or N Broadway in St. Louis (“St.” could be a suffix or part of a city name).

Traditionally, address elementalization has been implemented by building rule-based systems to handle each individual address format. This makes the process of building international address parsers very time consuming as one would need to implement a new rule-based parser for every single country. By building a statistical model to do this, implementing a new international format is as simple as training the model with a new data set.

Some interesting characteristics of postal addresses are that they have a grammar (as defined by the address format) and the elements are contextually dependent. These traits make the problem well-suited for natural language processing. There are natural language processing techniques that are used for similar purposes, namely part-of-speech taggers which are used to classify the parts of speech in a sentence.

For my final project, I looked at four different techniques of statistical part-of-speech tagging and applied them to the problem of postal address elementalization. The code is here.

The first strategy was a Hidden Markov model (HMM) tagger. HMMs are statistical models that can be used to find the most likely sequence of states for an input. This is done by using transition probabilities (the probability of a specific state given the previous state) and emission probabilities (the probability of the proposed state given the current input token). These probabilities are learned by observing a training set.  The model then tries to classify the input left to right, calculating the probability of a proposed state as the product of the transmission probability, emission probability, and maximum probability from the previous state. The model tries different state combinations over the input and returns the sequence of possible states with the highest overall probability.  A variation of this is the trigram HMM tagger, which uses the two previous probabilities to calculate the transition probability.

The second strategy was a Maximum-Entropy Markov model (MEMM) tagger. MEMMs are similar to HMMs in that they try to find the sequence of states that has the maximum total probability for an input. However, instead of just using observed counts for the transition and emission probabilities, we train a maximum-entropy distribution to calculate the probabilities. The main advantage of doing this is that we can have custom features factor into the probability, such as the current token length or whether the token contains a number. This allows for greater flexibility in tuning the tagger, but it makes the overall classification time slower.

The third strategy was a Transformation-Based Learning (TBL) tagger, also known as a Brill tagger. The TBL tagger generates rules based on observations in training. These rules indicate observed conditions on when a tag should be swapped with a different tag. During classification, initial tags are assigned to the terms based on the observed probability. The tagger iterates through the rules which were learned in training, swapping tags as necessary, until there are no more rules to be applied or a given score threshold is met.

The fourth strategy was a Conditional Random Field (CRF) tagger. CRFs implement a number of feature functions which take the proposed states and the input observation and return some value between 0 and 1. These features, like the features in MEMMs, can be just about anything, such as whether the current token is capitalized or whether it ends in -ed. Each of these feature functions has a different weight based on observations in training. The score for a sequence of states is calculated by summing the feature functions for every word over all words and then normalizing. Just like HMMs and MEMMs, we find the sequence that maximizes the overall probability.

From my initial work, I found that the Maximum-Entropy Markov model and the Conditional Random Field taggers consistently had the highest overall accuracy of the group. Both consistently had accuracies over 98%, even on partial addresses. For full format American addresses, these taggers had an accuracy around 99.7%. If I had to choose between them, I would go with the Conditional Random Field tagger as it was considerably faster in tagging the addresses.

I didn’t have enough time to finish everything that I wanted to implement, so this is still a work in progress. I still want to implement some smoothing techniques for unknown states, such as Katz backoff and Kneser-Ney interpolation. There are also a few more part-of-speech tagging techniques I would like to experiment with. I guess that old adage is true in this case – time flies when you are having fun…  🙂

Arduino and Motion

Today I took the Making Things Move With Arduino class taught by Ranjit Bhatnagar at NYC Resistor. I was particularly excited for this class as all of my Arduino projects have been rather stationary. This class was a great overview of some different ways to add motion to your Arduino project. It featured the Adafruit motor party pack.

We had five separate gizmos to play with: one DC motor, one stepper motor, two sizes of servos and one solenoid.

Left to right: small servo, large servo, solenoid, DC motor, stepper motor

We also assembled the Adafruit motor shield which simplifies using multiple motors with an Arduino. There was a lot of soldering to do up front but it was worth it! Once everything was assembled, we got to play with all of the different types of motors. We even got to keep the motors so the fun could go on all night long!

Arduino and Motors

What was my favorite part of the class? THE MOTOR PARTY!!! WOOOOH!!!

DIY Night Vision Camera

I found this Instructable about how to make your own night vision camera. It seemed to be a fun project, so I decided to give it a try.

The first step is to remove the infrared (IR) filter from the camera. I broke my first camera attempting to do this. I was far more careful with my second one and successfully removed the filter. The little blue chip is the IR filter:


This is how the photos look with the IR filter removed:

Kitty is not impressed.

After successfully removing the filter, the next step was to build an IR LED array to be used as a light for the camera. With a little bit of help, I was able to laser cut a perfect array of holes for the LEDs. Following the instructions, I assembled the LED array and turned it on, only to be disappointed by an incredibly dim light.

What went wrong? Here’s the point where I confess that I am relatively new to electronics, and so there are certain lessons that are yet to be learned.  I wired the LEDs incorrectly. I got a second batch and wired them together.


And still the array was too dim. It was time to really understand how the circuit of the array worked. There was something more at play here. I found a really cool LED array calculator online that helped me get to the bottom of my problem. I had to examine the LEDs more closely. The instructions use infrared LEDs from Radio Shack, which have a 940 nm wavelength, a 100 mA forward current and a 1.28 volt forward voltage. My first two attempts used LEDs that had a forward voltage of 1.5 volts, which meant that the LEDs were not getting enough power. I ordered a new set of IR LEDs with a lower forward voltage of 1.2 volts and assembled the array for the third time.


The third attempt was better. With the new array, I was able to capture photos in the dark!

A hand in the dark

 I did a little test to get a feel for how well the camera worked. First, I set up a small scene to photograph. I was interested to see how the camera could capture color and detail. Here is the control photo, taken with my normal camera:

Control Photo

First, I took a photo with the lights on. The room was somewhat dark, so the photo did not come out very clear:

Lights on

Next, I took a photo with the lights off, about one foot away from the objects. The detail was still somewhat clear, although differentiating colors was not really possible.


The second photo was taken from two feet away. At this point, some objects are no longer visible.

Two feet

The final photo was taken from three feet away. The objects are almost imperceptible at this point.


Although it was fun to build this, it isn’t very practical for real-world use. The major problem seems to be the power, as the 9 volt battery drains very quickly and is not strong enough to power many high-power infrared LEDs. If I go back to this project, the first step would be to build an array with a larger power supply and brighter LEDs. In the interim, I will just have to be content with taking nighttime pictures of things up close.


Recently, I’ve been spending a bit of time with Fritzing. It’s a piece of software that allows you to document prototypes, design circuits and manufacture PCBs. So far, I have only used the breadboard view to generate wiring diagrams for my Arduino and Sensors class, but I am still very impressed.

One of the circuits in my class involves wiring an accelerometer and a switch to an Arduino. Here was my attempt to photograph the circuit:

IMG_6845From the photograph, it’s not very clear on how to wire the accelerometer. There are too many wires in the photograph and it’s not easy to see where each wire terminates. Of course, I could try to recreate this circuit using wires with less slack, but there is a simpler solution – use Fritzing! The image generated by Fritzing makes it much easier to understand how the accelerometer is wired:

accelerometer_calIt’s fast and simple to generate the wiring diagrams. Fritzing has a bunch of predefined components (such as breadboards, switches, resistors and Arduinos) that you can drag and drop together. There’s also a great snap-to-grid functionality that ensures that components are connected. Fritzing allows you to import component libraries from other vendors so that you can prototype with correct representations of the components. For example, the accelerometer object comes from the Adafruit Fritzing library.

I’m really glad to see that there are excellent open-source tools available for this kind of thing. It makes sharing knowledge much easier. I’m already impressed by the breadboarding functionality, so I am looking forward to tinkering with the schematic and PCB views!