| Geological disasters in our country are characterized by high frequency,wide area and high harm,which have caused huge loss of life and property to the country.Therefore,safe and efficient post-disaster rescue is particularly necessary.At the same time,geological disasters are often accompanied by secondary accidents,which brings great danger to post-disaster rescue work.The application of intelligent self-rescue equipment can well solve this problem.In the rescue mission,effectively perceiving the surrounding environment is the first step for the self-rescue platform to perform rescue tasks.Therefore,it is particularly important to establish a perception algorithm to deal with the post-disaster unstructured road environment,so that the self-rescue platform can realize full-scene perception.In general,camera and Li DAR are the two main sensors for obtaining environmental information.The camera perceives the surrounding environment based on color and texture,which is greatly affected by light,and has a failure condition at night;As an active detection sensor,the Li DAR sensor can directly obtain high-precision 3D point cloud data of the surrounding environment and is not sensitive to light,so it is more suitable for post-disaster rescue scenarios.In order to ensure that self-rescue equipment can adapt to various lighting conditions for post-disaster rescue,this paper focuses on the use of Li DAR sensors to obtain point cloud data for scene understanding in unstructured road scenes.The main research contents are as follows:(1)First,because deep learning algorithms require a large amount of data,the semantic segmentation data of Li DAR point clouds with accurately labeled unstructured road scenes is scarce and the cost of manual annotation is huge.Therefore,this paper has sorted out a variety of computer simulation software that can be used for scene construction and data acquisition and based on the high degree of freedom and accuracy of the CARLA simulator,two typical unstructured road scenes in the field and dammed lake are selected.The scene is constructed,and the data of lidar point clouds of different resolutions and the corresponding precise semantic labels are obtained through simulation,which lays the foundation for the construction and implementation of the subsequent super-resolution network and semantic segmentation network.(2)Due to the sparse nature of point clouds,the number of point clouds on objects with small scales in the scene is small,which makes it difficult to capture the features of small-scale objects,and the problem of sample imbalance during training will also occur.Therefore,this paper combs the advantages and disadvantages of the interpolation algorithm and the deep learning algorithm,selects the deep learning algorithm with higher precision,and constructs a point cloud super-resolution network based on the projection map.For this reason,this paper introduces the MC-Dropout algorithm to estimate the distance distribution of point clouds,thereby reducing noise.Experiments show that the final super-resolution network can enhance the point cloud data collected by the simulator,and at the same time has better performance than the traditional interpolation algorithm,which lays a foundation for the follow-up research on unstructured road semantic segmentation.(3)Because the unstructured road scene has the characteristics of irregular road boundaries,and the large,medium,and small-scale objects in the scene must be effectively identified,according to the law of point cloud distribution,this paper constructs a point cloud-columnbased method.A 3D point cloud semantic segmentation network with a face voxel structure and a road boundary enhancement module,a multi-layer receptive field module and a data enhancement module.In the final experiment,while verifying the effectiveness of each module,Salsa Next and PVCNN are also selected for comparison,which shows that the network has stronger semantic segmentation ability while ensuring real-time performance. |
PDF Full Text Request