| Optical flow estimation is a fundamental task in the field of computer vision,which calculates the motion vectors of each pixel to obtain the motion information of objects in an image.It assists various advanced computer vision tasks.With the continuous development of computer technology,optical flow estimation technology has been widely used in fields such as autonomous driving,target tracking,unmanned aerial vehicle obstacle avoidance,and intelligent robots,and has become a research hotspot in the field of computer vision.In practical applications,most models do not consider the issue of platform differences,making it difficult to directly deploy deep learning-based optical flow estimation models on edge computing platforms with relatively limited computing resources.Therefore,lightweight optical flow estimation models still face many problems,such as a high number of iterations,difficulty in estimating occluded regions,and insufficient decoding capabilities of gated recurrent units.To effectively solve these problems,this paper mainly focuses on the development of lightweight optical flow estimation algorithms for edge computing platforms.The goal is to achieve high optical flow estimation accuracy on edge computing platforms with fewer computational resources and lower time costs.The main research contents of this paper are as follows:1.To address the problem of the existing optical flow estimation networks requiring a large number of iterations when using gated recurrent units as the iterative optimization basis,this study proposes a lightweight optical flow estimation algorithm based on convolution and axial attention.The algorithm employs Local Constraint and Local Dilation Module to replace self-attention with a combination of convolution and first-order axial attention,which focuses on different importance levels of the surrounding regions for each matching feature point with lower computational cost,generates more accurate correlation cost,and reduces the number of iterations of gated recurrent units,achieving the goal of lightweight.Additionally,Shuffling Convex Upsampling is proposed,which combines group convolution,shuffling operation,and Convex Upsampling to further improve accuracy while reducing the number of parameters.2.To address the issue of poor prediction of optical flow in occluded areas caused by reducing the number of iterations in research content 1,this study proposes a lightweight optical flow estimation algorithm based on multi-scale convolutional motion feature aggregation.The algorithm combines different scales of strip depthwise separable convolution and atrous depthwise separable convolution to focus on motion features within different ranges,replacing existing self-attention with global aggregation with lower computational cost to improve the estimation of optical flow in occluded areas.In addition,to address the lack of information interaction between matching features between consecutive frames in the Local Constraint and Local Dilation Module,Local Convolution and Shifted Windows Module is proposed.This module combines the advantages of convolution,self-attention,and cross-attention to simultaneously enhance the representation ability of matching features between consecutive frames and increase the interaction ability between the features of the two frames,further reducing the time consumption while improving accuracy.3.To address the issue of inadequate performance in large displacement optical flow optimization during each iteration caused by the limitations of using gated recurrent units in research topic 2,this study proposes a lightweight optical flow estimation algorithm based on iterative optimization using large convolution kernels.The algorithm replaces existing gated recurrent units with a large convolutional kernel optimizer,and using depthwise separable convolution and strip convolution with lower computational resources,can further improve the model’s accuracy without increasing computational time. |
