Robust Orientation Method Based on an Atmospheric Polarization Model for Hazy Weather

Accurate and reliable navigation data are key components in the Internet of Things (IoT). High-precision and stable autonomous orientation has attracted considerable attention regarding environments in which the global navigation satellite system signal is unreliable. Bionic polarized light orientation is a method based on the mechanism of insect navigation, and the orientation signal comes from nature, which has broad application prospects. However, there is significant dust or nitrogen oxide particle scattering and sunlight absorption in haze, reducing the polarization degree and abnormal polarization azimuth. This makes heading angle measurements with existing polarized light navigation and positioning methods inaccurate under hazy conditions. A polarization image denoising method based on deep learning can effectively eliminate the influence of haze on heading angles. Noisy polarization images can be decomposed into high- and low-frequency components containing different types of noise. Thus, this study proposes a polarization image denoising method that combines high- and low-frequency image decomposition with a multiscale two-level fusion strategy. This method decomposes the polarization image with noise into a set of low-frequency structural components and a set of high-frequency detailed components by utilizing multiscale Gaussian filters. On this basis, a multiscale two-level fusion strategy based on a complementary feature weighted fusion mechanism is also designed. In the high- and low-frequency merge and reconstruction module, after adding and fusing the high- and low-frequency enhanced features output by LF-UNet and HF-UNet, a residual convolution (Res-Conv) layer and an output convolution layer are used for residual prediction and then fused with the input image to acquire a high-quality polarization image. The proposed method is experimentally verified outdoors using a polarized light/inertial integrated heading measurement system and a vehicle. The experimental results indicate that the root mean-square error (RMSE) of the heading angle after denoising improved from 7.4931° to 0.2821°. This clearly demonstrates the effectiveness of the proposed method in enhancing orientation accuracy and promoting the application of polarized light navigation in hazy environments.