Research On 3D Point Cloud Target Recognition Algorithm In Autonomous Driving Scenario

In recent years,autonomous driving technology has gradually become the research focus of major unicorn enterprises and universities at home and abroad,among which 3D point cloud target recognition algorithm in autonomous driving scenarios is one of an important research frontiers.Druing the process of transforming from Level 2 assisted driving stage to Level 3,the automatic driving technology faces one key problem,which is that the current 3D point cloud target recognition algorithm has poor detection performance on small and medium-sized targets such as cyclists and pedestrians,and can not well detect targets in complex scenes.In order to solve these questions,the classic 3D point cloud target recognition algorithm PV-RCNN++ was used as the basic network framework to study the focused feature method of 3D point cloud.Two improvements were made,one was the fusion sampling based on D-FPS(Distance-Farthest Point Sampling)and F-FPS(Feature-Farthest Point Sampling),and another was to combine the CBAM attention mechanism module.By improving the recall rate of the network model for small and medium-sized targets,and enhancing the ability of the network model to capture the local spatial structure features of the target,three improved network models FS-PV-RCNN++,C-PV-RCNN++ and FSC-PV-RCNN++ were obtained.Finally,the ablation experiment based on the KITTI dataset was carried out,and the detection accuracy and detection effect diagrams of the improved network model and other network models were compared and analyzed.The results indicated that the FS-PV-RCNN++,C-PV-RCNN++ and FSC-PV-RCNN++ network models improved by the 3D point cloud focused feature methods,compared with the original network model PV-RCNN++,coul detect small and medium-sized objects with significantly improved accuracy.In conclusion,these methods can effectively enhance the ability of the 3D point cloud target recognition algorithm network to capture local spatial structure features under autonomous driving scenarios,strengthen the recognition effect of the algorithm network for small and medium-sized targets,and improve the detection performance of the algorithm network for complex scenes.