Online measurement of granular velocity of rotary drums by a fast PIV deep network FPN-FlowNet

Abstract

The measurement of velocity field of granules with respect to time in one rotary drum is of great significance in the study of mixing process. However, the existing measurement methods are not suitable to be used in online processing in real applications for the following reasons: limited spatial resolution, brightness sensitivity and long time-consumption. In this work, a hybrid framework containing an unsupervised deep network FPN-FlowNet and one particle-center detection model is proposed. FPN-FlowNet is built by introducing the FPN model for 4-level environment-robust feature extraction and weighted fusion for dense velocity field calculation; Particle center detection model is designed based on graph-cut and another well-trained deep model VGG16-LUnet to accurately obtain the particle positions. Extensive experiments including the actual granular flows in rotary drum along with the corresponding software simulation results demonstrate that our approach achieves competitive performance over the state-of-the-art method and satisfactory speed for online processing.