Abstract: As one of the important parameters of flow characteristics, flow velocity occupies an important position in the study of vertical pipeline lifting efficiency. To more accurately measure flow velocity and reveal the flow dynamics of vertical pipeline conveying systems, we focus on the solid–liquid two-phase flow in pipelines. In this paper, we study the method of pipeline velocity measurement and reveal the flow characteristics of the pipeline system. First, we use a high-speed camera to transform the flow velocity measurement into a computer vision problem, and combine the computer vision problem with deep learning technology to propose an A-RAFT (attention-based recurrent all-pairs field transforms) neural network model based on the attention mechanism. The model uses a convolutional layer to extract feature information and reduces the computational load through a pooling layer. Additionally, we introduce a correlation layer to perform inter-correlation operations on the feature information and calculate pixel displacement. In this process, the attention mechanism focuses on regions with flow velocity changes, enhancing the ability of the network to estimate velocity field variations. This helps the model better select and focus on key features in the input data, providing more accurate feature information for matching. Consequently, the estimation accuracy of the model is improved, particularly for the boundary regions of solid particles in solid–liquid two-phase flow. The model also effectively estimates flow rates for particles of varying shapes and sizes, with enhanced overall performance and accuracy. In addition, this paper constructs a combined real and virtual dataset for training the neural network model. The dataset is based on nine types of classical single-phase flow field data, and real particle texture information is fused into the dataset through real experiments to enhance data diversity. This dataset effectively simulates the optical flow changes of the pixels in the front and back frames in real experiments. The proposed model is evaluated with this dataset, and the results show that the model achieves high-precision velocity field computation on synthetic images, and the estimation error is 15.6% lower than those of other existing models. In the simulation experiments of solid particle transportation in vertical pipelines, the proposed model demonstrates accurate estimation performance on the collected real flow field data, with relative errors of lower than 5% for the measurement of particle velocities. These errors are derived from comparisons with the true values. The results validate the method in terms of both estimation accuracy and the generalization ability of the model. This study can provide new insights for solid–liquid two-phase flow characterization in energy extraction, tunneling, wastewater treatment, and long-distance pipeline transportation.