Research On 3D Scene Semantic Segmentation Algorithm Based On Deep Learning

In computer vision,3D semantic segmentation is an important basis for 3D scene understanding and analysis.It is widely used in the field of 3D perception such as unmanned driving,autonomous robot and augmented reality,etc.Improving the accuracy of semantic segmentation is of great significance for semantic understanding and analysis.At present,the performance of the 3D semantic segmentation algorithm based on deep learning is much better than the traditional segmentation algorithm,but the accuracy can not fully meet the actual needs,there is still a lot of room for improvement.In order to improve the accuracy of 3D semantic segmentation,this paper studies the 3D semantic segmentation algorithm based on deep learning,the specific research content is as follows:PointNet is one of the advanced algorithms in 3D semantic segmentation of point clouds.This paper improves on PointNet to improve the accuracy of semantic segmentation.Aiming at the problem that PointNet does not use the local structure of the point cloud,which leads to its low ability to identify fine-grained objects,this paper proposes a multi-scale local feature extraction module,which searches the local neighborhood of the point cloud through the KNN algorithm,learns the contribution of each features in the neighborhood to local features,and obtains the local feature by weighted summation,and connects the multi-scale local feature to obtain the multi-scale local feature.Aiming at the problem that PointNet treats all feature channels indiscriminately,which limits the network's ability to express,this paper uses the channel attention enhancement module to adaptively learn the importance of different channel features,increase the weight of important channel features and suppress the unimportant channel features.Finally,the global feature and local feature of multiple levels are connected for semantic segmentation,and the weighted loss function is used to assist the training.Compared with the PointNet,the oAcc and mIoU of the algorithm in this paper are improved by 7% and 9.8% respectively,and the objects with complex spatial structure and local segmentation details are better identified.In order to further apply more and more widely RGBD images to 3D semantic segmentation,this paper studies the 3D semantic segmentation algorithm based on RGBD multi-frame fusion,and improves the effect of 3D semantic segmentation after multi-frame fusion by improving the accuracy of image semantic segmentation.First,this paper proposes a RGBD semantic segmentation algorithm based on dual data streams.Aiming at the problem that the depth information is not fully utilized in image semantic segmentation,two ResNet networks are used to extract RGB features and depth features respectively,and the multi-modal feature fusion module is designed to adaptively select important parts of RGB features and depth features to fully fusion,reducing the influence of RGB images on factors such as lighting and occlusion.Aiming at the problem that the segmentation result of deep network is rough due to the down sampling,the detail and accuracy of the segmentation result are improved by the integration of low-resolution and high-resolution features step by step.Then,the results of multi-frame RGBD semantic segmentation are fused by bayes' rule to realize 3D semantic segmentation.Compared with ResNet network,the oAcc and mIoU of the two-stream semantic segmentation algorithm in this paper are improved by 4.81% and 7.69% respectively.Moreover,it is not easily affected by the error of texture information,and the segmentation details should be clearer.A good 3D semantic segmentation results can be obtained by accurate multi-frame fusion of image semantic segmentation.Experiments show that the 3D semantic segmentation algorithm based on point cloud is more accurate than the PointNet.In the 3D semantic segmentation algorithm of RGBD multi-frame fusion in this paper,the accuracy of RGBD semantic segmentation is good,and through effective fusion,the 3D semantic segmentation results are obtained,which verifies the feasibility and effectiveness of the algorithm.In order to collect 3D point cloud and RGBD images,this paper designs and implements 3D point cloud scene acquisition system and RGBD image acquisition system,and elaborate the scheme from two aspects of hardware and software.Finally,summarize the work and determine the next research direction.