自監督使用單光達生成多光達物件點雲視角

Abstract

使用光達點雲的三維物件偵測被廣泛的應用於自駕車以及路側系統的物件偵測與追蹤。然而使用單光達點雲進行偵測的表現會受到物件遮蔽的影響。許多過往的研究提出將物件點雲補齊來提升三維物件偵測的表現。多數研究的模型輸入為單一物件的點雲，需要額外的前處理將物件點雲從光達掃描中取出。有些研究則是以整個光達掃描點雲作為模型輸入，但仍是需要有物件的三維標記，而新場景的標記的取得耗時且困難。本研究利用路側系統的多光達掃描之特性，設計了一個自主標記的流程。該流程透過去背景之多光達融合點雲，使用DBSCAN演算法得到物件三維標記，並以此在多光達融合點雲上取得物件之多光達視角。補點模型的訓練則是將多光達點雲拆成單一光達點雲作為輸入，並以物件之多光達視角點雲作為訓練的輸出。此流程使得補點模型得以在沒有人工標記下完成補點，在體積像素(Voxel)IoU相對於沒補點前增加約17%，體積像素(Voxel) recall則是相對於沒補點前提升約22%。 DBSCAN的3D AP@IoU=0.25在經過補點後提升約20%，並超越使用雙光達作為輸入的結果。深度學系模型的偵測表現亦有提升，如:PointPillar及PV-RCNN在KITTI資料集中Easy類別的3DAP@IoU=0.5分別提升約3%和1%。

The 3D object detection using point clouds scanned by LiDARs has been widely used in applications such as self-driving vehicles and roadside vehicle detection and tracking. However, the detection performances using singleLiDAR-scanned point clouds suffer from occlusion. Several previous works proposed to complete the point cloud to improve 3D detection. Most works take the input of the object point clouds, which requires additional steps for processing the LiDAR-scanned point clouds. Some works complete object point clouds by taking the input of full LiDAR-scanned point clouds, but they require human-labeled 3D object bounding boxes, which are difficult and costly to obtain for new scenarios. This work designed a self-labeling method that exploits the characteristics of multiple LiDAR-scanned point clouds collected on roadside units. The self-labeling method labels the objects with DBSCAN using the fused point cloud scanned by multiple LiDARs that have been background-removed. The labels are then used to extract object points from the fused point clouds. The point completion model is trained using the single-LiDAR-scanned point cloud as input and using the extracted object points from the fused point cloud as the training target. This labeling method allows the point completion model to be trained without humanlabeled bounding boxes. The voxel IoU of the point-completed point clouds increased by 17% compared to that of the raw point clouds. The voxel recall of the point-completed point clouds increased by 22% compared to that of the raw point clouds. Additionally, the detection performance of DBSCAN, i.e., 3D AP@IoU=0.25 increased by 20% after applying the point completion algorithm. The deep-learning-based models also benefit from applying the algorithm. The 3D AP@IoU=0.5 for the easy targets in the KITTI dataset of PointPillar and PV-RCNN increased by about 3% and about 1%, respectively.