Research On Unsupervised Video Moving Object Segmentation Algorithm Based On Dual-stream Feature Fusio

As a key first step in intelligent video analysis,video moving object segmentation is the premise and foundation of higher-level image understanding tasks,and has been one of the hot research fields in recent years.Since there is no need to manually label information,unsupervised video moving object segmentation has higher practical value and more challenges.On the one hand,to solve the problems of objects’ appearance change and insufficient use of temporal features,we propose a method for unsupervised video moving object segmentation based on visual attention mechanism and temporal memory mechanism.On the other hand,to solve the problem of scale change,a method for unsupervised video moving object segmentation based on multi-scale features is proposed.Extensive experiments on two challenging public benchmarks(i.e.DAVIS 2016 and Seg Track V2)show the effectiveness of our methods.The specific work is as follows:1)An unsupervised video moving object segmentation method based on temporal memory mechanism and appearance attention mechanism is proposed.Firstly,the motion features and appearance features of moving objects in video sequences are extracted by two-stream network,and the moving objects are segmsed by two-stream fusion features.According to the characteristics of the moving target segmentation task,the mechanism of integrating channel attention,spatial attention and global attention is introduced to learn the appearance features,which effectively strengthens the feature learning of the target region and reduces the interference of background noise.Then,a fast memory unit Conv SRU is designed to cache the historical location and appearance features of the target,and then combine the current features of the target,making full use of the temporal features to refine the segmentation results without increasing the computational complexity too much.At the same time,we propose a consistent loss function,which further reduces the influence of noise,and can achieve a better segmentation effect even when the target is temporarily stationary.2)Aiming at the problem of target scale variation in the process of video moving object segmentation,an unsupervised video moving object segmentation method based on multi-scale features is designed.Firstly,an encoder with cross structure is designed to integrate temporal and spatial features,and the object motion is integrated into the appearance learning in a hierarchical way,which fully considers the deep interaction information between the appearance features and the motion features,and improves the effective representation of the temporal and spatial features of the moving target.Then,a multi-scale feature extraction module is introduced to extract the spatio-temporal features of the cross-encoder output,which improves the robustness of the model to scale changes.Finally,a feature correction module is designed,which can fuse deep features,shallow features and multi-scale features,so as to obtain a relatively fine segmentation result.