End-to-End Multi-Task 3D Object Detection Method Based on Bird Eye·s View Images
Due to spatial limitations during the scanning process of LiDAR during autonomous driving,the collect-ed point cloud data loses too much information.Based on compressed representation,a novel multitask end to end 3D object detection model in Bird's Eye View is presented.Firstly,texture feature extractor and semantic feature extractor were used to fuse the high-level and low-level feature layers of a feature pyramid network(FPN)to obtain an extend-ed feature pyramid layer,which eliminated the coupling of medium target(cyclist)and small target(pedestrian)at the same level of feature pyramid layer during the auto-driving process,improving the regional detail of the point cloud topology information.Secondly,the Focal loss and CIOU loss were introduced in the loss function to improve the pro-portion of negative sample and limit the bounding box.As a result,the bounding box can converge more accurately in the regression process.Finally,The experimental results show that the proposed method has better detection ability and higher detection accuracy in small and medium targets detection of auto-driving.
NETL
NSTL
万方数据
