Deep Learning Based Semantic Segmentation Of Point Cloud For Autonomous Driving

LiDAR point cloud semantic segmentation means assigning a meaningful label to each point in the LiDAR point cloud.LiDAR can provide point cloud data containing high-precision environment information for automatic driving systems.Environmental Semantic information can be obtained by semantic segmentation of point cloud data,which can assist automatic driving systems in decision-making and path planning.Autonomous driving systems require real-time,accurate,and robust perception to the surrounding environment.Traditional point cloud semantic segmentation methods mostly rely on manual feature descriptions,which are not only time-consuming but also have low generalization capabilities.The deep learning method gradually shows its unique advantages in automatic feature extraction,but most of the 3D deep learning methods have limited real-time running capabilities in dealing with massive point clouds.Therefore,many works exploit the image based semantic segmentation method to carry out research on semantic segmentation of point cloud.However,these methods have not achieved the best trade-off between real time performance and segmentation accuracy.And the conversion of 3D point cloud data into two-dimensional representations may lead to important information loss,which affects the segmentation performance of the model.In response to these problems,we focus and do research on the following in this paper:(1)In order to make a better balance between the real time performance and accuracy of point cloud semantic segmentation,this paper proposes a residual attention based real-time semantic segmentation for point cloud.The model uses the spherical projection method to project the original point cloud into a distance image as the network input.The model uses the residual unit of ResNet-34 to form the basic segmentation network with the encoder-decoder structure.And channel attention module is introduced in the residual unit to build a residual attention module to strengthen the network’s attention to important channel information.At the same time,combined with the context awareness module,the receptive field of network feature extraction is improved,and the skip-connection is used to make up for the detailed information in the process of semantic recovery.Through experiments,the model has realized real-time operation,and the point cloud segmentation accuracy has been greatly improved compared with existing methods.(2)In order to alleviate the problem of information loss caused by projecting point cloud data into range image,this paper proposes a point cloud semantic segmentation framework based on the fusion of different modals.The framework can be divided into two stages.First,3D features are extracted by the 3D feature extraction module based on the neighborhood from each point,which is effectively fused with the distance image after the point cloud projection.With improved residual attention based real-time semantic segmentation,additional input channel Mask,Focal loss,the framework processes the fusion features to obtain the point-by-point classification predictions.The experimental results verify that the fusion of 3D features from point cloud and projection images can effectively improve the segmentation performance of the model,indicating the effectiveness of the framework.