MPA-RAFT: Motion Perception Attention for Particle Image Velocimetry

Particle image velocimetry (PIV) is a widely used method for calculating velocity fields from continuous pairs of particle images, enabling the nonintrusive estimation of complex particle motion fields and playing a crucial role in the visualization analysis of fluid mechanics. Although many studies have applied deep learning (DL) techniques to address fluid motion estimation issues, some existing DL methods encounter problems of erroneous estimation and insufficient accuracy when dealing with more complex flow fields. Aimed at improving PIV estimation accuracy, this article proposes a PIV motion estimation method based on a recurrent update mechanism, named motion perception attention recurrent all-pairs field transform (MPA-RAFT). Many methods do not consider the integration of motion information across different dimensions and directions. Therefore, we introduce a novel motion perception attention (MPA) module specifically designed for PIV tasks, which perceives and aggregates motion information from different dimensions to enhance the capture of motion information. We analyzed and validated the effectiveness of the MPA-RAFT model on the dataset, and experimental results indicate that MPA-RAFT demonstrates strong generalization ability and performs well in handling complex flow fields. Compared with the RAFT, it has achieved a significant average improvement of 30% on multiple flow types, while attaining state-of-the-art performance on PIV tasks, surpassing previous DL and traditional methods. Furthermore, the results from the experimental jet flow images show that MPA-RAFT can handle real world experimental scenario flows effectively.