Research On LiDAR Semantic Segmentation Algorithm Based On Deep Learning

With the vigorous development of 5G technology,computer vision and 3D sensor technology,autonomous driving technology has gradually gained more and more attention and applications.The perception system in autonomous driving mainly uses sensor technologies such as Li DAR and camera as the "eyes" of autonomous driving,so that it plays an important role in recognition and assistance in autonomous driving.The semantic segmentation of natural images is one of the most important technologies for extracting information from sensor data.Although some breakthroughs have been made in the semantic segmentation of natural images in recent years,there are still problems such as inaccurate target edge segmentation and low overall segmentation accuracy.This paper mainly focuses on camera-based real-time semantic segmentation and Li DAR semantic segmentation technology.For solving the above difficult problems,corresponding improvements are made in terms of network structure design and loss function to improve the segmentation accuracy.The main research contents of this thesis are as follows:1.A real-time semantic segmentation method based on boundary feature aggregation and multi-scale fusion(BAMFNet)is proposed.Aiming at solving the problems of lack of multi-scale information interaction and inaccurate object edge segmentation in the real-time image semantic segmentation,a real-time semantic segmentation network that takes into account multi-scale information fusion and introduces boundary features to improve segmentation accuracy is designed and implemented.The multi-scale fusion module is not limited to the hierarchical multi-scale feature fusion in the current popular network structure,but can achieve more fine-grained multi-scale feature fusion.It can effectively increase the receptive field of each network layer and help the network learn more detailed information.At the same time,a boundary feature aggregation module is added in the middle part of the network,which realizes the guided learning of intermediate representations by using edge information,and it can achieve more effective extraction of target edge features,thereby improving the overall segmentation accuracy of the network.By conducting comparative experiments on the Cityscapes dataset and visualizing the segmentation results of each model,the experimental results prove that the proposed method achieves better performance in meeting the requirements of real-time performance compared with the state-of-the-art methods.2.A point cloud semantic segmentation method based on boundary feature refinement(BFR-Net)is proposed.Aiming at handling the problem that the state-of-the-art point cloud semantic segmentation method SCF-Net cannot well segment the spatial and geometric boundary regions between different classes,a boundary feature refinement model based on SCF-Net is designed.The model adds a boundary feature refinement module on the SCF-Net,which finds the k nearest neighbor points for each point,and judges whether the neighbor point and the center point belong to the same category.For the neighborhood points and center point belonging to the same category,we use the weighted sum of neighborhood points features to enhance the feature of the point;for the neighborhood points and center point belonging to different categories,we use symbiotic relationship between the center point and their neighborhood points to enhance the feature of the center point.Through the experimental verification and comparison on the indoor point cloud dataset S3 DIS,the effectiveness and feasibility of the proposed method are proved,and this method can obtain better segmentation results.3.A point cloud semantic segmentation method based on variable weight cross-entropy loss function is proposed.In order to solve the problem of unbalanced class distribution of Li DAR point cloud data in the real world,a variable weight cross-entropy loss function(EXPCELoss)is designed for point cloud semantic segmentation.The variable weight cross-entropy loss function aims to assign a larger weight to the class with a small number of samples,and slightly reduce the weight of the class with a large number of samples,so that the network can strengthen the feature learning of the target with a small number of samples.In addition,the variable-weight cross-entropy loss function is simple to implement and can be plug-and-play.It is applied to several point cloud semantic segmentation algorithms with the best performance at present,and a comparative experiment is carried out on the indoor point cloud dataset S3 DIS and the urban point cloud dataset Sensat Urban.The experimental results show that the variable weight cross entropy loss function can be applied to multiple point cloud methods and can improve the segmentation effect of the corresponding algorithm on two point cloud datasets,thus proving the plug-and-play,effectiveness and robustness of the variable weight cross-entropy loss function.