Optical Flow Estimation via Weighted Guided Filter with Nonlocal Steering Kernel

The attractive advantages of the weighted guided filter (WGF) in the optical flow estimation method are that WGF can alleviate the influence of outliers and preserve the motion boundaries of moving objects. However, the WGF-based optical flow estimation still suffers from the problem of halo artifacts near the motion boundaries because of motion occlusion and illumination changes under adverse outdoor conditions. In this chapter, we introduce a novel optical flow estimation method based on the weighted guided filter with nonlocal steering kernel (WGF-NLSK). The proposed approach combines the edge-preserving capabilities of the WGF-NLSK's ability to capture global motion patterns, enabling accurate flow estimation in challenging scenarios. The WGF ensures that motion boundaries are preserved while smoothing the optical flow field, and the NLSK improves robustness to large displacements, occlusions, and complex motion structures. A key innovation of the WGF-NLSK method is its ability to handle fine details and subtle motions, making it particularly effective in scenarios with intricate structures and large variations in motion. We employ a coarse-to-fine framework to improve computational efficiency and allow the method to manage large motions and fine structures at multiple scales. Through careful parameter selection, we optimize the method for different datasets, achieving a balance between accuracy and smoothness in the flow field. Extensive experiments on benchmark datasets, including Middlebury, MPI-Sintel, and the Wilted Plant Leaves dataset, demonstrate the superior performance of our method compared to traditional optical flow algorithms. The results highlight the method's ability to capture fine details, preserve motion boundaries, and accurately estimate optical flow in complex scenes. Our approach shows particular promise in scenarios involving large displacements and fine motion boundaries, outperforming existing techniques in these domains. The chapter concludes with a discussion of future directions, including improvements in computational efficiency and potential applications in real-time video analysis and dynamic motion tracking.