Abstract: [Purposes] Pole-like objects along roads are key elements that constitute the urban road skeleton and semantic scene.Their precise instance extraction is crucial for high-definition map production,autonomous driving environment perception,and smart city management.However,current extraction methods relying on a single data source struggle to achieve a balance between geometric,attributive,and semantic information.Furthermore,multi-source data fusion methods are plagued by issues such as data registration accuracy,density inconsistency,and temporal mismatch,making them unable to meet the requirements for component-level extraction of road furniture in urban scenarios. [Methods] To address the aforementioned problems,this paper proposes a method for instance extraction of road pole-like objects by fusing vehicle-borne images and point clouds.Firstly,high-precision registration is used to align image textures with point cloud geometric structures.Secondly,an initial clustering based on frustum bounding boxes generated from the point cloud is employed to separate potential pole-like objects.Then,precise extraction is performed by combining the spatial structural features of pole-like objects.Finally,principal component analysis is introduced to segment adhering or adjacent pole-like object clusters into instances. [Findings] Test results from certain road sections in Nansha district,Guangzhou,show that both recall and precision exceed 90%. [Conclusions] This method fully leverages the advantages of multi-source data,significantly improving the completeness of pole extraction and the accuracy of instance segmentation in complex urban scenes.It holds important value for enhancing the automation level of 3D understanding in road scenes.

Key words: MLS point cloud, imagery, instance segmentation, pole-like objects extraction, feature extraction