Research On Moving Object Segmentation Algorithms Based On Deep Learning

Moving object segmentation is a critical task in many computer vision applications as it aims to accurately segment all moving objects in the current scene.At present,there are still many problems in moving object segmentation,such as co-planar and co-linear motion degeneration,background misclassification,motion noise,and multi-scale.To effectively solve the above problems,this paper proposes three moving object segmentation methods that combine geometric methods and deep learning techniques to improve the results of moving object segmentation.The main research content of this article is as follows:1.Geometric analysis based motion-aware architecture for moving object segmentation.Compared with traditional methods,deep learning-based moving object segmentation methods demonstrate great advantages.However,many CNN-based methods only rely on motion information from 2D optical flow maps to segment moving objects,they usually ignore the incorrect 2D optical flow in optical flow maps and do not consider the problems of coplanar and collinear motion degeneracy.In order to address difficulties caused by motion degeneracy,this paper proposes the geometric analysis based motion-aware architecture for moving object segmentation.For coplanar and collinear motion degeneracy,this paper proposes the reprojection cost and optical flow contrast cost,which adopt geometric methods to calculate the difference between the motion of moving objects and that of the background.Meanwhile,to detect moving objects with weak 2D motion features,this paper combines the 2D motion and3 D motion of objects to strengthen motion features,and designs the bidirectional motion constraint.In order to reduce background misclassification,this paper introduces the motion detection branch to generate motion heatmaps and capture the motion region.To accurately obtain the mask of moving objects,this paper adopts the motion instance segmentation branch to extract high-level motion cues and segment moving objects.In addition,extensive experiments conducted on moving object segmentation datasets demonstrate the effectiveness of our method.2.Appearance fusion based motion-aware architecture for moving object segmentation.While the reprojection cost and optical flow contrast cost can be used to detect objects with coplanar and collinear motion,there are still some issues that need to be addressed.The reprojection cost is difficult to segment moving objects far away from the camera,and the optical flow contrast cost cannot segment moving objects whose 2D flow magnitudes are consistent with the background but have different directions.Thus,this paper adopts the background 2D optical flow to adjust the cost values of moving objects at different distances,and propose the balanced reprojection cost.For the optical flow contrast cost,this paper adds the 2D motion direction information of objects to designs the relative optical flow contrast cost.Moreover,to effectively segment non-rigid moving objects and reduce the background misclassification,this paper introduces the appearance fusion based motion-aware architecture for moving object segmentation.In this architecture,multi-modality co-attention gate is used to promote the adaptive fusion between appearance and motion information of objects.To reduce the motion noise in low-level motion features,this paper designs a multi-layer motion fusion module which aggregates motion features at different levels and enriches the semantic information of low-level motion features.To prevent background misclassification,a motionbased attention module is proposed to utilize the motion information to suppress salient static objects and highlight moving objects.Finally,this paper conducts many experiments on common moving object segmentation datasets to demonstrate the effectiveness of our proposed method.3.Adaptive feature fusion based motion-aware architecture for moving object segmentation.To effectively address the motion noise and multi-scale issues in moving object segmentation,this paper proposes the adaptive feature fusion based motion-aware architecture for moving object segmentation.To obtain richer motion features and improve the accuracy of moving object segmentation,the cost-based moving feature generation module integrates the reprojection cost and balanced reprojection cost to generate the optical flow difference cost.The bidirectional optical flow contrast cost is designed to detect slowly moving objects.In order to minimize the effect of motion noise on segmentation results,this paper improves the multilayer motion fusion module,and proposes the adaptive multi-layer fusion branch to promote the adaptive fusion of motion features at different levels,which aims to enhance the attention to moving features with important levels.For multi-scale problems,this paper adopts the layerby-layer decoder and position mask prediction module,by introducing the multilayer supervision strategy and combining the advantages of features at different levels,the network could better segment moving objects of different scales.Besides,our proposed method has been shown to be superior through extensive experiments.