We present an approach to automatic visible light/infrared (VL/IR) image registration that leverages multiple visible light apertures for fast computation on resource-constrained systems. VL/IR registration is computationally challenging due to the different modalities of image generation. Although feature-based algorithms for direct registration exist, these methods proved too complex to reliably perform registration on low-cost, embedded processors in real time. We instead employed a second VL camera to dynamically estimate 2D translations aligning the brightest (warmest) objects in the IR video stream with their counterparts in the first VL video stream. Regions of interest are first selected based on the brightest areas in the IR image, as our application is primarily concerned with detecting objects warmer than background. The same broad region - e.g. the lower-left quadrant of the frame - is then selected in the VL1 and VL2 images. The translation that best registers the first VL ROI to the second is estimated through template matching. Because all apertures in our camera system are fixed and coplanar relative to one another, the translation that best aligns the IR ROI to the VL1 ROI can be estimated from the translation from the VL2 ROI to the VL1 ROI. This approach provides dynamic registration of 1080P video at upwards of 10Hz on an ODROID-XU4 single-board computer, while also allowing the processor time to render the IR-augmented video stream at 20Hz. Imagery collected using Deep Analytics’ IR Boom Camera will be presented to demonstrate the approach.
|