Research On Visual Content And Scene Understanding Based On LiDAR Point Cloud

Lidar sensors have become essential devices in most fields such as autonomous driving and robotics.Compared to vision sensors,they capture more accurate depth information and are more robust under various lighting conditions.Semantic segmentation of LiDAR point clouds is an important task in autonomous driving technology,as it provides fine-grained scene understanding and is complementary to object detection.For example,semantic segmentation can help self-driving cars distinguish between drivable and non-drivable road surfaces,and can reason about road surface functions,such as parking lots and sidewalks,which cannot be replaced by other algorithms in autonomous driving systems.However,annotation of large-scale 3D data is notoriously cumbersome and costly.As an alternative,weakly-supervised learning alleviates such a need by reducing the annotation by several order of magnitudes.We first project the point cloud onto spherical coordinates,so that the point cloud becomes a more compact form of expression,so that the point cloud data can perceive the contextual information of the scene in a more efficient way.In addition,we self-train the network to propagate limited labels to unlabeled pixels.Finally,we de-noise the pseudo-labels generated during training using class prototypes and prediction uncertainty to alleviate the problem of network misfitting to noisy data.To achieve efficient online class prototyping update,we further reduce the noise in pseudo-labels through a momentum update-based network and exploit prediction uncertainty and a symmetric cross-entropy function.Since contrastive learning requires rich and diverse examples as keys and anchors,we propose a prototype memory bank capturing class-wise global dataset information efficiently into a small number of prototypes acting as keys.An entropy-driven sampling technique then allows us to select good pixels from predictions as anchors.Experiments using a light-weight projection-based backbone show we outperform baselines on three challenging real-world outdoor datasets,working with as low as 0.001%annotations.