HawkEye

Events from a conventional sensor

HawkEye emits only the moving edges in a scene, as address-event representation data. Once the static scene is gone, a target a few pixels across separates from the clutter that would otherwise bury it. HawkEye enables operators and detection processes to identify information that might otherwise be missed.

Send what moved, not the whole scene

Below is a park path at dusk. The original recording has limited contrast and dynamic range so it required preprocessing with our EagleEye algorithm.

The trees, the grass and the path are standing still, so they produce almost no events and almost nothing is sent for them. What leaves the device is the walker and the dog, and the data rate rises and falls with them.

0:00

Deploy resources selectively

Events can provide cues to higher-cost downstream processing. A detector or classifier that would otherwise sweep a whole frame can run on the region that moved. You can wait for preliminary evidence before risking the deployment or potential exposure tied to active sensors.

Example

Below is drone footage at sunset. A small target crosses left to right in the glare beside the sun, almost impossible to notice without HawkEye. Watch the original first and see if you notice anything in the sky.

0:00

Bats at dusk

0:00

Timing, position, and polarity

HawkEye takes in video and sends out an event stream of when, where, and which direction each motion event occurred and nothing else. Static objects generate no data. The bandwidth required to monitor is significantly reduced, so what leaves the device is proportional to what actually moved.

In the drone clip's source video, HawkEye's event stream compressed with gzip is 4.2 Mbps on average, dropping to 64 kbps when the drone stops moving.

HawkEye's event generation stage is configurable, enabling its output to be compatible with any event-based cameras and algorithms you already use. Events can also be filtered by frequency, which helps in cluttered scenes and when the camera itself is moving.

HawkEye's events compressed with gzip, against the drone clip as visually lossless H.264.

FPGA implementation

Form: AXI Stream IP block
Target: Xilinx Ultrascale+ Zynq
Added power¹: <100 mW
FPGA resources: <1%
Output: AER over network or to file
Temporal resolution: sensor frame rate

¹ Added to the EagleEye IP block on the same FPGA.

Try it on your own footage

Contact us to arrange a demonstration based on a video you provide. We have a range of mechanisms for facilitating this, depending on the nature and sensitivity of your data.

Contact