Research On 3D Scene Semantic Segmentation Technology Based On RGBD Camera

The robot has the ability to learn and understand its external environment,which is an important prerequisite for interacting with the external environment to complete specified tasks.In understanding the external environment and learning,semantic segmentation is one of the key technologies.This paper studies the deep neural network semantic segmentation technology based on RGBD camera to achieve indoor scene semantic segmentation.The RGBD camera is the basic technology for making RGBD data sets,and has played a role in promoting the establishment and development of sample data sets.The research design of the model in this paper is mainly carried out in two steps.The first step is to conduct a lightweight model study on the RGB images in the data set to achieve preliminary segmentation.The second step is to integrate depth information on the basis of the model to achieve optimal segmentation.In view of the contradiction between the complexity and accuracy of deep neural networks,a lightweight and effective backbone network is selected for feature extraction,and a lightweight semantic segmentation model is designed.The model uses an encoder and decoder structure as a whole,adopts parallel sampling and introduces jump connections,To achieve the combination of different feature information;zero-centralization of the image and data enhancement,speed up the convergence speed,and improve the model performance.In view of the diversity of indoor scenes and the occlusion relationship between objects,the overall structure and technology of the lightweight network model are retained,a branch structure is added to introduce depth information,and a semantic segmentation model incorporating depth images is designed to complete the optimization of the model.The model training was carried out on the SUN RGBD indoor scene data set,and the lightweight network model and the semantic segmentation model fused depth image were evaluated and compared.After the fusion of depth images,the accuracy of the model has been improved,the pixel accuracy rate has been increased from 71.6% to73.3%,the average pixel accuracy rate has been improved from 46.1% to 47.9%,and the regional merge ratio has been improved from 31.3% to 33.2%.The experimental results verify the effectiveness of this method.