基于多方向特征融合的室外三维目标检测方法

Outdoor 3 D Object Detection Method Based on Multi-Direction Features Fusion

三维目标检测方法是自动驾驶环境感知系统的重要技术之一,但是现有三维目标检测方法大多存在目标位置识别精度不足、目标朝向预测偏差较大的问题。为此,提出一种基于多方向特征融合的三维目标检测方法。对场景点云进行数据编码,建模距离、夹角等信息并转换成伪图像。利用所提多方向特征融合骨干网络进行特征提取和融合,基于多方向融合特征,利用以中心为基准的检测器进行潜在目标的回归和预测。该方法通过点间距离、夹角建模提供点间的相互关系,丰富特征信息,利用多方向特征融合增强骨干网络的特征提取能力,从而提高位置及朝向的预测精度。实验结果表明,该方法在KITTI和nuScenes数据集上的平均精度均值分别为64.28%和50.17%,相比次优方法分别提升了0.36和1.30个百分点,且在朝向预测准确率对比实验中均取得了最好的平均朝向相似度和平均朝向误差。此外,泛化性实验结果验证了所提多方向特征融合骨干网络能提高网络检测能力并大幅减少参数量,从而提升检测方法的应用表现。

The 3D object detection method is one of the significant technologies in the environmental perception of autonomous driving.However,most existing 3D object detection methods have the problem of inaccurate position recognition and large orientation deviation.To address these issues,a 3D object detection method based on multi-direction features fusion is proposed.First,to perform data encoding for a point cloud scenario,modeling distance and angle and transforming into pseudo image.Second,a multi-direction feature-fusion backbone is proposed for features extraction and fusion.Finally,based on the fused features,a center-based detector regresses and predicts potential objects.Distance-angle modeling can supply the relationship between points and enrich features.The multi-direction feature-fusion backbone enhances the ability of features extraction and improves the accuracy of position and orientation estimation.The experimental results show that the mean Average Precision(mAP)of this method on the KITTI and nuScenes datasets was 64.28%and 50.17%,respectively,which is an improvement of 0.36 and 1.30 percentage points,respectively,compared to those of the suboptimal method.In addition,the best Average Orientation Similarity(AOS)and mean Average Orientation Error(mAOE)were achieved in the orientation prediction accuracy comparison experiments.The generalization experimental results verified that the proposed multi-direction feature-fusion backbone network can improve network detection ability and significantly reduce the number of parameters,thereby improving the application performance of the detection method.