Unfolding Hyperspectral Imagery…
We’re happiest when we’re experimenting; trying new things to see what works and what breaks. For the most part, we can use the huge amounts of open source data available online thanks to hardworking, generous people around the world. But sometimes to try new ideas, we need new data, or even new hardware.
Recently, we’ve been looking at the power of hyperspectral imagery to provide information beyond the ordinary RGB. We’re working with the great team over at Living Optics and using their new dev kit to see what we can find. In order to test this though , we needed data, so we spent three days in the field recording. From warehouses to stairwells, offices to living rooms, we’ve been pointing cameras at a range of sights.
Top left: Sony RGB camera. Top right: InfiRay T2 Pro longwave infrared (LWIR) camera. Bottom: Living Optics hyperspectral imager (HSI). We’re still working on our setup!
Obviously it rained heavily which curtailed our plans, but we still gathered some valuable data that we’re now going through and testing some of our initial ideas. Even though we’re focusing on the hyperspectral side now, having a range of simulataenous (albeit not boresighted) data collects will allow us to fuse the data in novel ways.
Top right: RGB. Top left: LWIR. Bottom left: Hyperspectral scene view. Bottom right: 1 - the spectral angle (i.e. similarity) from Tom’s teal coat.
In our initial testing, we were able to pick someone out (Tom shown here) consistently across a range of scenes using different lenses and even through glass. This level of consistency should allow us to provide some level of object tracking over time, and potentially add support to a re-identification decision.
All images show 1 - the spectral angle (i.e. the similarity) between the HSI pixel spectrum and the target spectrum, with brighter being a closer match. Top left: the probe image. Top right: an image from later on in the same video. Bottom left: a different video with Tom inside a building behind glass. Bottom right: a different lens (8mm instead of 35mm), showing Tom as two bright pixels on the right.
We’ll be capturing more soon, so if there’s anything you want to see tested, then please get in touch! We’d love to give back to the open source data community, so watch this space…