Sub-pixel matching method for low-resolution thermal stereo images

摘要: In the context of a localization and tracking application, we developed stereo vision system based on cheap low-resolution 80×60 pixels thermal cameras. We proposed threefold sub-pixel matching framework (called ST for Subpixel Thermal): 1) robust features extraction method phase congruency, 2) rough these in pixel precision, 3) refined accuracy local coherence. performed experiments our very images (acquired using manufactured) as high-resolution from benchmark dataset. Even if congruency computation time is high, it was able to extract two times more than state-of-the-art methods such ORB or SURF. modified version correlation applied feature space matching. Using simulated stereo, investigated how con-gruency threshold sub-image size can influence accuracy. then proved that given setup resolution images, being wrong 1 leads ±500 mm error Z position point. Finally, showed could four matches baseline + OpenCV KNN images. Moreover, were robust. More precisely, when projecting points standing person, got standard deviation ≈ 300 gave 1000 mm.