Authors: Dr. Federica Bianco
contact: fb55 @
Prof. S. Koonin, Dr C. Mydlar, Mr. M Sharma


  • Resources (this research is reproducible)

  • presentation:

  • ACM BuildSys 2016

  • Preprint



We image an NYC lightscape at 4Hz while running a shutter at 119.75Hz to generate a 0.25Hz beat frequency of the original 120Hz flicker of lights. The flicker is the rectified 60Hz oscillation of the input AC current. Phase changes in the flicker reflect changes in the load of the unit's circuit (e.g. apartment, changes occur when appliances are turned on and off). Phase changes in the original 120Hz frequency (or in the 60Hz input current), appear as phase changes in the beat frequency.

Detect windows in an NYC night lightscape.

NYC lightscape (hpertemporal image stack)

Flux distribution in NYC lightscape pixels

Windows and urban light in NYC lightscape pixels

Extract the lightcurves and analyze them

Above, a gif movie of a window with detectable 0.25Hz oscillations

Below, the location of the window with detectable flicker (gif above) in the lightscape, its time series (lightcurve) and the power spectrum of the time series, with the 0.25Hz frequency indicated by a red line. We detect frequencies near our 0.25Hz target. The small discrepancy (0.02Hz) is well within the accuracy of our instrumentation and of the grid frequency delivery.

The time series set is decomposed with PCA

The first and second components are sine waves in qudrature: they reconstruct the flicker signal in which we are interested.

PCA decomposition of a 25 seconds imaging bursts at 4Hz

Above: lightcurves projection on the plane of the first and second principal components. Lightcurves at a large radius on this plane display a strong flicker component in their time series. Because not all light technologies flicker at 120Hz (LED lights and modern fluorescent lights do not) not all lightcurves show the signal. Lights in red are selected for analysis

Histogram of the distance of the lightcurve projection on the plane of the first and second principal components, and the two components.

Below are the PCA selected lightcurves with the sine fit to the lightcurve overplotted, with a phase calculated as arctangent(PC1/PC2) and the power spectrum. In the gif the time series are ordered by decreasing signal-to-noise ratio. Some of the selected time series are shown further below.

Frequency and phase distribution of the time series

The time series cluster in phase by building, but all phases are seen in the scene.

Identified phases colorcoded by phase

Urban lights flicker phase evolution

The following plots show the phases and phase changes in the lights that we selected in 5x5 min bursts of imaging separated by 15 minutes.

Phase evolution: phases relative to a reference-light phase are shown in their evolution within a 5 minute imaging butsts (5 time stamps separated by 37.5 seconds) and between bursts separated by 15 minutes. Colored lines indicate shifts larger than our sensitivity (error bar at the bottom left) and the color coding reflects the buildings distribution: line in the same color indicate windows in the same building.

Pairwise phase offset evolution at 15 minutes time intervals. lights are ordered by building so the block structure reflects building-wide changes.

Read our paper!