Road-Object Instance Segmentation Methods Of Point Cloud

Autonomous driving is the intersection of intelligent transportation and artificial intelligence technology.It aims to realize unmanned intelligent vehicle driving,so it needs to combine multiple artificial intelligence technologies.The earliest technology used in automatic driving systems was object detection technology based on two-dimensional images.However,the coarse-grained inference task cannot accurately predict the specific shape of the object in the application,and the limitation of camera images brings more uncertainty.Therefore,exploring the robust 3D point cloud data format combined with more fine-grained instance segmentation inference tasks is of great significance to the research of autonomous driving technology.The research content of this report is point cloud-based instance segmentation methods in autonomous driving scenarios.Instance segmentation is the refined inference of instance-level semantic segmentation of the target.It is necessary to further distinguish different instances of objects based on location after pixel-level semantic segmentation,which can accurately depict the edge contours of different instance objects in the scene.The existing point cloud instance segmentation method is realized by projecting a point cloud scene on a two-dimensional image in combination with deep learning methods.Although some progress has been made with such methods,there are still some problems.First,the outdoor scene has sparse point clouds with complex and huge background noise,which leads to extremely low target saliency.It is difficult to directly perform fine-grained inference tasks,and the front view projection may cause aliasing and occlusion between objects in the same axial direction.In addition,there are a lot of difficult samples in point cloud instance segmentation,resulting in the average accuracy of point cloud instance segmentation is still at a low level.This report researches on the above issues,specifically:(1)This thesis first proposes an instance segmentation model of point cloud based on pillar projection.It presents a two-stage point cloud instance segmentation framework,which combines point cloud object detection with semantic segmentation to achieve refined point cloud instance segmentation.In order to enhance the saliency of point cloud objects,this report first uses an object detection algorithm to make a preliminary understanding of the scene.Filtering out most background noise by preserving the pillar of point cloud of interest,which is convenient for refined instance segmentation.Encoding and spherical projection of the extracted pillar to regularize the point cloud.Since the pillar contains at most one object,the projection of pillars can avoid the aliasing of objects.Then,a lightweight two-dimensional convolutional neural network is used to achieve efficient and fine-grained instance segmentation.The twostage model based on pillar projection in this report realizes high-quality instance segmentation by combining the location information and semantic information of the object.Experiments on public datasets indicate that the performance of the proposed algorithm can exceed the state-of-the-art.(2)This thesis further researches the instance segmentation of difficult point cloud samples,and proposes a three-dimensional point-based instance segmentation model,which effectively improves the recognition and segmentation of difficult point cloud samples.In order to enhance the ability to express the features of difficult samples,this thesis proposes an instance segmentation network that directly extracts 3D features from the original point cloud.Extracting the refined semantic structure features of objects through the point-wised network can avoid the loss of topology details of objects brought by projection.Specifically,first,filter the noise through the ROI extraction module with cross-validation to improve the salience of the target.At the same time,negative samples with the same distribution are introduced to enhance the generalization of the segmentation network.Then,segment point clouds with a multi-class focus loss function network module.The network refines the recognition and segmentation of the points within the pillar,and focuses on mining features of difficult samples.Experimental results confirm the effectiveness of the proposed method,which can enhance the performance of the difficult samples in instance segmentation from point clouds.The two point cloud instance segmentation models proposed in this report both exceed the current international best algorithms in experimental comparison.According to their different characteristics,the two algorithms have different applicability.The instance segmentation model based on the pillar projection can achieve efficient and lightweight instance segmentation in point cloud scenes with few difficult samples.The other model based on three-dimensional points achieves more accurate point cloud instance segmentation through refined feature extraction and higher computational cost.