Sky Region Recognition and Polarization Orientation via Multi-Scale Gradient Connected-Component Extraction for a Bio-Inspired Sensor

Sky-region recognition using a bio-inspired polarization sensor helps reduce occlusion-induced errors in polarization orientation and supports applications such as Global Navigation Satellite System (GNSS) non-line-of-sight error mitigation and atmospheric irradiance estimation. However, retaining usable sky measurements while rejecting non-sky interference remains challenging in scenes with extensive occlusions. This paper proposes a sky-region recognition and polarization orientation method that combines multi-scale grayscale gradients with connected-component extraction. Multi-scale gradients are computed from the grayscale image derived from the polarization measurements. Gradient, intensity, and degree of polarization (DOP) thresholds are then combined to obtain an initial binary sky mask. Connected-component analysis is then performed, and the largest component is retained as the final mask, which is used to select the polarization measurements for heading estimation. Experimental results on the unmanned platform in scenes containing building, vegetation, and pedestrian occlusions show that the proposed algorithm achieves a heading root-mean-square error (RMSE) of 0.48°, corresponding to a 36.0% reduction relative to the DOP-gradient-based method.