HU-FlowNet: A Method for Trajectory Detection in Dense Granular Flows Integrating the Hungarian Algorithm, U-Net, and UnLiteFlowNet

The accurate tracking of particle trajectories in dense granular flows is vital for optimizing industrial processes involving rotary drums. However, traditional tracking methods suffer from poor stability and low accuracy under complex conditions. To address these issues, this paper proposes HU-FlowNet, a trajectory detection method integrating U-Net for particle localization, UnLiteFlowNet for velocity field estimation, and the Hungarian algorithm for global position matching. Specifically, UnLiteFlowNet is used to predict particle positions at frame <i>i</i> + 1 based on their locations at frame <i>i</i>, while U-Net detects the observed particle positions in frame <i>i</i> + 1. The Hungarian algorithm is then employed to establish optimal correspondences between the predicted and observed positions, thereby determining the updated locations of particles and achieving accurate tracking across frames. Experiments on dense particle flow in rotating drums demonstrate that HU-FlowNet outperforms conventional trackers (Euclidean, CSRT, Boosting, MedianFlow, and simple online and realtime tracking), achieving a root mean square error of 4.854, mean absolute error of 1.913, and a trajectory coverage rate of 94%. The proposed method demonstrates strong robustness across varying operating conditions, validating its effectiveness and generalization capability and enables stable, continuous tracking of hundreds of particles in complex, dense environments.