0228011-1第60卷第2期/2023年1月/激光与光电子学进展研究论文融合语义和激光点云空间可见性特征的3D行人检测熊璐，邓振文，田炜*，王之昂同济大学汽车学院，上海201804摘要车载激光雷达为智能驾驶汽车提供精确的周围空间几何信息而成为车载主流传感器。为克服单传感器对目标检测的局限性，对激光点云的几何特征、空间可见性特征和图像语义信息在合理设计的网络框架中进行融合，进一步提升3D行人检测精度。首先采用高效的三维空间光线投射算法形成空间可见性特征编码；其次融合图像语义类别信息，增强点云特征；最后定量和定性分析各附加信息和相关超参数对检测结果的影响。实验结果表明：相比单帧点云，结合历史前10帧点云后3D行人检测精度提升32.63个百分点；进一步融合图像语义和点云空间可见性信息，相比基准方法，所提方法的检测精度提升2.42个百分点，且超过部分主流方法，更加适用于交通场景的3D行人检测。关键词目标检测；图像与点云融合；点云空间可见性；智能驾驶环境感知中图分类号TP391.4文献标志码ADOI：10.3788/LOP220712Three-DimensionalPedestrianDetectionbyFusingImageSemanticsandPointCloudSpatialVisibilityFeaturesXiongLu,DengZhenwen,TianWei*,WangZhiangSchoolofAutomotiveStudies,TongjiUniversity,Shanghai201804,ChinaAbstractVehicularlightdetectionandranging(LiDAR)hasbecomeastandardsensorinautomotivebyofferingaccurategeometricinformationofthesurroundingregionforintelligentdrivingvehicles.Inordertoovercomethelimitedperformanceofasinglesensorforobjectdetection,thegeometricandspatialvisibilityfeaturesofLiDARpointcloudsarefusedwithimagesemanticinformationinanetworkframeworktoachieveaccuratethreedimensional(3D)pedestriandetection.First,aneffective3Dray-castingalgorithmisintroducedtoproducespatialvisibilityfeatureencodings.Second,theimagesemanticinformationisincorporatedtoimprovepointcloudfeatures.Finally,theimpactofaddedinformationandrelatedhyperparametersondetectionfindingsarequantitativelyandqualitativelyexamined.Experimentalfindingsdemonstratethatcomparedwiththesingleframepointcloud,the3Dpedestriandetectionaccuracyisenhancedby32.63percentagepointsafteraggregatingthelast10framesofthepointcloudinhistory.Byfurtherfusingimagesemanticsa...
