Robust NLTV ‐ L 1 Optical Flow Estimation Under Adverse Outdoor Conditions

It is known that flow field estimation is a low-level task of computer vision, and robust estimation of optical flow is essential for achieving precise motion segmentation. This chapter introduces a robust nonlocal total variational (RNLTV- L 1 ) approach for optical flow estimation, aimed at accurately computing flow fields in challenging outdoor environments. Furthermore, the computed flow field is utilized for performing motion segmentation. Conventional approaches to estimating optical flow and segmenting motion typically emphasize image sequences captured in ideal conditions, paying little attention to challenging environments such as haze or fog. The RNLTV- L 1 framework, however, maintains edge integrity and reduces excessive smoothing around motion boundaries even in complex outdoor situations. For more reliable motion segmentation, the method employs motion-pattern-driven background–foreground separation, enabling precise detection and delineation of moving objects. Existing optical flow estimation and motion segmentation techniques have traditionally been tailored for normal conditions, often overlooking environmental factors like haze, rain, or fog that significantly degrade performance. Our method offers a robust solution by integrating a geodesic-based weighting scheme and leveraging non-local information to improve edge preservation and maintain accuracy under adverse conditions. Additionally, the application of our method to motion segmentation further validates its efficiency in real-world scenarios, where clear delineation of objects from the background is essential. This chapter further highlights the novel contributions of the RNLTV-L 1 optical flow method: the introduction of a support-weight function to avoid blurred motion boundaries, a coarse-to-fine optimization strategy, and its validation on multiple challenging datasets. A series of extensive experiments are performed to evaluate the method's robustness across datasets including MPI-Sintel, Middlebury, Foggy Zurich, Traffic, and CDnet2014. These datasets encompass various environmental challenges, ensuring that the method's effectiveness extends beyond controlled environments.