基于数据特征的多传感器融合实时目标检测

Multi-sensor fusion real-time target detection based on data characteristics

为了进一步降低目标检测出现的误检率,提出了一种基于传感器数据特征的融合目标检测算法。首先,为了减少部分离群噪声点对点云表达准确性的影响,采用统计滤波器对激光雷达原始点云进行滤波处理;其次,为了解决点云地面分割在坡度变化时,固定阈值会导致分割不理想的问题,提出了自适应坡度阈值的地面分割算法;然后,建立KD(k-dimensional)树索引,加速DBSCAN(density-based spatial clustering of applications with noise)点云聚类,基于Andrew最小凸包算法,拟合最小边界矩形,生成目标三维边界框,完成聚类后的目标点云位姿估计;最后,将激光雷达检测到的三维目标点云投影到图像上,投影边界框与图像检测的目标边界框通过IoU关联匹配,提出基于决策级的三维激光雷达与视觉图像信息融合算法。使用KITTI数据集进行的测试实验表明,提出的点云聚类平均耗时降低至173 ms,相比传统的欧氏距离聚类,准确性提升6%。搭建硬件实验平台,基于实测数据的实验结果表明,提出的融合算法在目标误检率上比YOLO v4网络降低了约10%。

This paper proposed a fusion object detection algorithm based on sensor data features to further reduce the false detection rate of target detection.Firstly,it used statistical filters to filter the original LiDAR point cloud to reduce the influence of some outlier noise points on the accuracy of point cloud expression.Then,it proposed a ground segmentation algorithm with adaptive slope threshold to solve the problem that fixed threshold would lead to unsatisfactory segmentation when the slope of point cloud ground segmentation changed.Finally,it established a KD(k-dimensional)tree index.It projected the LiDAR-detected 3D target point cloud into the image and match the projection bounding box and the image detection target bounding box by IoU association.It proposed a decision-level-based 3D LiDAR and visual image information fusion algorithm.The test results using the KITTI dataset show that the proposed fusion algorithm reduces the average time spent on point cloud clustering to 173 ms,which is 6%more accurate than the traditional Euclidean distance clustering.The experimental results based on the mea-sured data show that the proposed fusion algorithm reduces the target false detection rate by about 10%compared to the YOLO v4 network.