Research On Optical Flow Algorithm Based On Cost Volume Enhancement

Optical flow is the motion of image points between adjacent frames in video,and optical flow algorithms play an important role in video motion analysis.Variational methods dominate the early optical flow algorithms,which solve optical flow by minimizing energy functional.In recent years,with the development of deep learning technology,optical flow algorithms have begun to focus on deep learning methods.Optical flow calculation model based on deep learning consists of three parts: the first part is the feature encoder,which is used to extract matching features of input images;the second part is the cost volume calculation module,which obtains cost volume by performing correlation operation on matching features;the third part is the optical flow decoder,which solves optical flow from cost volume.Cost volume is the core component of optical flow calculation model,it stores the matching cost of image points between adjacent frames,high-quality cost volume can improve the accuracy of optical flow results,making the calculated optical flow more valuable.This paper conducts in-depth research on optical flow algorithms based on deep learning,and finds that the matching features extracted by previous algorithms are noisy and contain almost no semantic information,moreover,there are a lot of outliers in original cost volume directly calculated using matching features,which will interfere with subsequent optical flow calculation and affect the accuracy of optical flow results.In order to solve these problems and improve the accuracy of optical flow results,this paper proposes an optical flow algorithm based on cost volume enhancement.First,this paper builds a context feature encoder to extract context features of reference frame and target frame,the context feature encoder consists of multiple convolutional layers and it uses skip connections,which can ensure that the encoder does not degenerate and facilitate the transfer of information between different layers,moreover,channel attention mechanism is introduced to adjust features of different channels to improve their overall representation ability.The features extracted by context feature encoder are no longer noisy and contain rich semantic information.Secondly,this paper designs a cost volume correction module,which generates a weight for each matching cost in original cost volume by combining deep learning method and mathematical method,and then uses these weights to correct original cost volume to reduce outliers.Finally,in order to make full use of semantic information in context features,this paper proposes an information fusion module,which uses context features to calculate semantic information volume,and then integrates it into corrected cost volume to obtain final cost volume.Compared with original cost volume,the outliers of final cost volume are greatly reduced,and the quality of final cost volume is obviously enhanced.In addition,this paper also discusses the optical flow initialization methods and the checkerboard artifacts in optical flow results.This paper quantitatively compares the optical flow algorithm based on cost volume enhancement with other algorithms,and proves the effectiveness of proposed algorithm by comparing experimental results.