| The foreground object is the most attractive part of an image,generally the first place the human eye observes,and it is also the main research object of various image analysis.Image segmentation and extraction of foreground objects is one of the key steps in tasks such as scene analysis,object detection,autonomous driving,and metaverse construction.The segmentation results will directly affect the final presentation effect.Therefore,correctly segmenting foreground objects in images has important engineering and practical significance.Because of its importance and high research and development difficulty,image segmentation technology has always been a research hotspot in the field of computer vision.As researchers continue to invest,various methods emerge each year,but most of them only deal with color images.Since the color image cannot provide the field three-dimensional space information,when the color information between the foreground object and the background is similar,the foreground information will be difficult to be extracted correctly.In recent years,depth cameras have gradually entered everyone’s field of vision,and the corresponding technologies have also been rapidly developed.The acquisition of RGBD images has become more and more convenient,and this direction has gradually become a new research hotspot in the field of image segmentation.This paper proposes a multi-foreground object segmentation method based on RGBD images.By using depth information to make up for the lack of color information,it can achieve more accurate extraction of foreground objects.The main work of the thesis includes:1.This thesis proposes a novel road plane pixel extraction method.The existing road surface extraction algorithms generally extract the entire road surface area,but it is difficult to extract the area that is in direct contact with the foreground object.The algorithm proposed in this thesis can effectively disconnect the depth connection between the pavement and the foreground objects by fitting the pavement depth information and using the information difference between the original depth map and the fitted pavement to extract the pavement pixels.2.This thesis proposes an adaptive multi-threshold depth image stratification algorithm.Firstly,the data set required by the experiment is made according to the relative position of foreground objects in the depth map.Then,the ResNet(Residual Network)image classification model is trained based on the data set,so that it can correctly predict the layers of the depth image.Finally,the prediction result is used as the input of the multi-threshold Otsu method to obtain the deep layer and reduce the manual interaction.3.The color image is segmented using the Mean Shift algorithm,resulting in oversegmented color blocks.We define a standard for region merging,including depth layer constraints and area threshold constraints,and merge the over-segmented foreground objects to extract foreground objects.In order to verify the accuracy of the algorithm proposed in this paper,this paper conducts experimental comparisons on datasets in three different scenarios,and compares them with four excellent image segmentation algorithms.The experimental results show that the algorithm has high accuracy in various indoor and outdoor scenarios. |
