This non-rotating necklace uses sensors to determine its orientation with respect to magnetic North and illuminates four coloured blobs (on a circular LED strip) at the four cardinal points: North, East, South, and West. The coloured blobs thus stay more-or-less stable as the wearer rotates - interestingly contrary to one's usual expectation that the two would rotate together. It's intended for dancers - aiming to accentuate dance movement, be moderately eye-catching, and be intriguing, but not to be excessively obtrusive.
Additionally:
- each coloured blob slowly changes colour, following a random walk in a region of colour space;
- the light intensity is scaled with the wearer's rotation speed, becoming brighter if they spin;
- each blob is actually a (mild) colour gradient, not a uniform colour, and the direction of the gradient flips depending on whether the wearer is rotating clockwise or anticlockwise;
- the sensor data and reconstructed angle can be sent by short-range radio to a receiver, to be displayed live or recorded on a laptop.
The sensors measure the current linear acceleration, gyroscope rotation rate, and magnetic field, each in three axes.
We assume the wearer will be largely vertical, with the necklace not necessarily horizontal, but more-or-less stable in the reference frame of the wearer's shoulders. In normal use, as the wearer moves, the acceleration vector is dominated by gravity. To estimate which way is up, in the device reference frame, we take a smoothed (low-pass-filtered) version of the acceleration data. We project the gyroscope and magnetic field data onto this horizontal plane, which also greatly simplifies the mathematics and computation needed, compared to the general 3d orientation estimation problem.
The magnetic field data gives an absolute orientation with respect to the Earth's magnetic field, but with a very noisy signal. On the other hand, the gyroscope gives reasonably good data about the rotation rate, which one can integrate to get an orientation estimate, but it will tend to drift with time as errors accumulate. We fuse the two with a complementary filter, choosing the time constant empirically to balance reducing jitter from the magnetometer noise against giving a fast response to quick rotation. Additionally, when the angular rotation rate is small (less than two pixels/second), we smooth the resulting orientation estimate.
The result is responsive and repeatable. It's not always completely linear with respect to the actual North angle, as one can see in the second video above, but in practice it's pretty good, as one can see in the first video above.
Sample graphs from a double spin are below.
Calibrating the sensors is important. The magnetometer is calibrating by finding the maximum and minimum reachable values for each axis, using a 30-sample rolling mean (probably this could be improved, and poor calibration may well be responsible for the non-linearity of the result). The gyro zero offset for each axis is automatically re-calibrated whenever the device seems to be at rest (comparing all the raw sensor values against a rolling window). If this is within the first 20s of startup, it blinks eight LEDs white. The gyro max range is set to 1000 degrees/s (less was not enough for fast spins), and the accelerometer range to 2g.
Meanwhile, each cardinal-point region of LEDs has a base colour from a random walk within a fixed RGB cuboid. Each random walk has a current position, velocity, and duration (in cycles). When the duration reaches 0, a new random velocity is chosen. The walk bounces off the internal faces of the RGB cuboid. Each region is three pixels wide, anti-aliased onto four pixels of the strip. One end is the base colour, and subsequent pixels are shifted by a factor of the current pixel colour velocity - this gives each region a slowly changing colour and a slowly changing colour gradient. The gradient direction is flipped based on the direction of rotation (to give a mild visual highlight to changes of rotation direction), slightly offset from zero to avoid flicker at rest. The regions have brightness scaled by the angular velocity (above 0.2 rad/s).
The overall brightness is good for normal or dim artificial lighting; clearly visible but not dazzling, and well-diffused without the individual pixel LED points being obtrusive. In overcast outdoor light, it's visible but rather subdued.
The refresh rate, for a main loop including reading the sensor data, computing the orientation estimate, and refreshing the display, is a comfortable 50 Hz. Reading the magnetometer data takes a long time (around 21ms) and is not needed at high frequency, so is only done once every four cycles. Printing data to the serial link slows this down considerably, while sending data over the radio link has little effect.
Power is from a Fenix ARB-L16-700U 3.6v 700mAh Li-ion battery, which conveniently has a built-in micro-USB charging socket - so the necklace can connected either to the battery or to the Pro Micro micro-USB socket (but not both at once!). Power to the neopixel strip is taken from the Pro Micro regulated power, which will be less efficient than the battery supply, but makes it conveniently comparable when connected to the 5v micro-USB supply. Adafruit estimate around 20mA per full-brightness single-colour pixel, at 5v, but the measured power here is much less, around 8mA per full-scale single-colour pixel. Running all pixels full-power would still vastly exceed the Pro Micro regulator (1730mA vs 500mA max), and also give tiny battery life - but that would also be uncomfortably bright for non-daytime use. Measured power, running on the battery, is around 65mA before the neopixels start up, 85mA steady state, and up to 220mA when rotating. That should give battery life of 3 to 8 hours. Power consumption is not notably affected by whether the radio is used.
One can see the sensor and orientation estimate data simply by printing it to the Arduino serial link, but it's much more convenient to be able to do this wirelessly. The nRF24L01P+ transceiver gives a cheap and easy way to do this, sending 32-byte packets to a similar receiver attached to another Arduino, itself connected via USB to a laptop.
Visualising the data is essential for development, and also interesting to see the sensor properties and the quantitivae parameters of movement in use. The client software is a simple application that reads the incoming data from a serial device (either directly connected to the necklace USB, or indirectly via the radio and a receiving Arduino), and uses gnuplot to draw various graphs dynamically. It can also record and replay data streams.
The main Arduino libraries used are MPU6050 for the MPU9150, Adafruit_NeoPixel-master, RF24, and I2Cdev. The first included Madgewick and Mahoney orientation estimation code, but they seemed to work less well (for this constrained application) than the scheme above.
The electronics consists of:
- a Sparkfun Pro Micro 3.3v/8MHz Arduino processor (much cheaper clones are available);
- an MPU-9150 board, containing a three-axis accelerometer, gyroscope, and magnetometer;
- an nRF24L01P+ 2.4Gz Transceiver board, for short-range radio;
- half an Adafruit NeoPixel Digital RGB LED Strip 144 LED - 1m White, ie 72 pixels; and
- a Fenix ARB-L16-700U 3.6v 700mAh Li-ion battery, with built-in micro-USB charging socket.
The first three are wired up so they fold into a tight package (insulated with tape):
and then that and the battery are cable-tied to the necklace structure:
See connections.md for the wiring and parts.md for a list of all the parts.
The necklace structure is a 500mm length of Stainless Steel Knitted Wire Mesh Tubing - Hop Filter Mesh - 22mm Diameter (from The Mesh Company), with the ends reinforced with some 0.5 and 1.0mm stainless wire, and a hook from 2mm stainless wire. The neopixel strip is just loose within this, insulated at the open end with clear tape, and held with a cable tie on its connecting wires at one end.
The diffusing shell is a piece of silk satin, nominally white but actually cream coloured, cut from a 0.2m length from John Lewis (506 10321 barcode 2986 9855). Overall 580mm long, 130mm wide for 95mm, 97mm wide for the main part, with a diagonal intermediate part, with fray-check solution applied to the edges before cutting. Hand-sewn into a tube. We tried various other diffusers, including other John Lewis fabrics and Rosco quiet light grid cloth (E462) (which looks promising but is only sold in long rolls); this is a good balance between diffusion and light loss, and feels good on the skin.