Research On Bubble Flow Simulation And Object Tracking Algorithm Optimization

In the study of gas-liquid two-phase flow,the trajectory information of moving bubbles in liquid is an important manifestation of bubble dynamic behavior.Obtaining a set of reliable trajectory data is the foundation for subsequent bubble dynamic characteristics research.Compared with other objects,the bubble flow data studied in this paper is small and dense,and is prone to deformation and aggregation during movement.Based on the characteristics of bubble flow data,this thesis improves the YOLOv5 object detection algorithm and combines it with the optimized Deep Sort object tracking algorithm to perform real-time trajectory tracking of moving bubbles.The main content includes the following three aspects:Firstly,in response to the shortcomings of YOLOv5 algorithm in detecting small and dense objects,a coordinate attention mechanism is added to the backbone network to enable the algorithm to pay attention to interested objects in a larger area.During the object tracking process,the tracking information predicted by the Deep Sort algorithm will be correlated with the detection results of the YOLOv5 algorithm.Therefore,enhancing the ability of YOLOv5 algorithm to detect bubble flow data can improve the integrity and accuracy of trajectory tracking results.The experiment shows that the introduction of coordinate attention mechanism improves the detection accuracy of the improved model on the same test set.Secondly,in response to the cumbersome problem of collecting the real trajectory information of bubbles,a bubble flow simulation algorithm based on generative adversarial networks is proposed.The real trajectory information of bubbles is an important basis for evaluating the performance of tracking algorithms.This thesis uses the generative adversarial networks to obtain rich bubble images.Through the random walk algorithm and the bilinear interpolation algorithm,the movement path and shape change of bubbles are designed respectively.Then,combined with the dynamic behavior of bubbles and the statistical results of measured data,multiple groups of bubble flow data with known trajectory information are simulated.This can not only efficiently obtain the required experimental data,but also achieve the goal of enhancing the robustness and universality of tracking algorithm.Finally,based on the discrete Fréchet distance,this thesis proposes a trajectory similarity measure to optimize the performance of the object tracking algorithm.By calculating the similarity between the real trajectory and tracking trajectory of simulated bubble flow,feedback results are obtained,and training and tracking parameters are continuously adjusted to improve the tracking effect of the model on bubble flow data.Among all training models,we choose the algorithm model that is more suitable for bubble flow data to track the measured data,in order to obtain reliable trajectory information.