水下视觉SLAM的图像滤波除尘与特征增强算法

An Image Dust-Filtering and Feature Enhancement Algorithm for Underwater Visual SLAM

将视觉SLAM(同步定位与地图创建)方法应用于水下环境时,扬起的沉积物会导致SLAM特征点提取与追踪困难,而且人工光源的光照不均匀还会引起特征点分布不均与数量较少。针对这些问题,设计了一种水下图像半均值滤波除尘与光照均衡化特征增强算法;根据水中杂质的像素特征,按照“检测-滤波”的顺序采取从外至内的半均值滤波过程消除扬起的沉积物在图像内造成的干扰;同时,通过统计光照均匀、充足区域内的像素分布,得到同一地形下不同位置处的环境特征相似的规律,并将其用于求解水下光照模型,将图像还原为光照均衡的状态,以此来增强图像的特征,进而实现更多有效特征点的提取。最后,利用该滤波与增强算法对多种海底地形数据集进行处理,并在ORB-SLAM3算法下测试运行。结果表明,滤波与增强后的数据集能够将特征点提取数量和构建地图的点云数量平均提高200%。综上,图像滤波除尘与特征增强算法能够有效提高视觉SLAM算法的运行效果与稳定性。

When the visual SLAM(simultaneous localization and mapping)method is applied to underwater environment,the interference caused by raised sediments makes it difficult to extract and track SLAM feature points,and the uneven illumination by artificial light sources causes the uneven distribution and small number of feature points.To solve those problems,a semi-mean dust-filtering and illumination equalization based feature enhancement algorithm is designed for underwater images.According to the pixel characteristics of the impurities in water,the semi-mean filter algorithm removes the raised sediment in the image from outside to inside in the order of detection-filtering.And,the distribution of pixels in areas with sufficient and even illumination is counted,and a law is obtained that the environmental characteristics at different locations in the same terrain is similar.Based on the law,the underwater illumination model is solved to restore the raw image into an image with even illumination,and thus image features are enhanced to extract more effective feature points.Various underwater terrain datasets are processed by the filtering and enhancement algorithm,and some tests are carried out with the ORB-SLAM3 algorithm.The results show that the number of feature points extracted and the number of point clouds for mapping are increased by 200%in average by using the filtered and enhanced datasets.So,the image dust-filtering and feature enhancement algorithm can effectively improve the performance and stability of visual SLAM algorithms.