Improved RefineDNet algorithm for precise environmental perception of autonomous earthmoving machinery under haze and fugitive dust conditions

Abstract

Non-homogeneous haze and fugitive dust are frequently encountered at earthwork construction sites, which can lead to errors and omissions in vision-based perception modules of autonomous earthmoving machinery. This study proposes an improved RefineDNet dehazing and fugitive dust removal algorithm to solve this problem. Compared with the original RefineDNet, the improved algorithm integrates feature attention and attention gate modules to address the non-homogeneous distribution of haze and fugitive dust, and integrates a Res2Net module to process multiscale objects. The proposed algorithm effectively enhances the visibility and realness of hazy or dusty images, thereby improving the object detection accuracy significantly. The peak signal-to-noise ratio and structure similarity index measure achieved by the proposed method were on average 2.42 and 0.0699 higher, respectively, than those achieved by state-of-the-art algorithms on a dataset of haze and fugitive dust at a construction site. The object detection accuracy of the restored images was 0.7160, which is an improvement of 45.6% compared to that of the original photographs in the on-site tests.