融合点柱网络和DETR的三维复杂道路目标检测

3D complex road target detection method by fusing PointPillar network and DETR

三维目标检测是智能驾驶的关键技术之一,但是存在处理数据量大、预设特征参数多等问题。针对三维特征选择设置与实际的目标特征相关性较差的问题,提出了一种融合点柱网络和DETR的检测方法。首先,利用支柱编码来降低冗余点云的影响,使特征提取的匹配性更强,且提升了计算效率;其次,基于DETR解码器的预测模块,使用多头注意力机制建立全局特征与预测集的关联映射,并行计算出相关性最强的预测结果,避免了人工依赖先验知识介入参数导致的不确定性;最后,在公开数据集上进行了验证,相较于原点柱网络,平均检测精度均值提升了19.14%,FPS提升了3,与其他典型算法相比也有较大的提升。

Three-dimensional target detection is one of the key technologies for intelligent driving,but problems like large amount of processed data and many preset feature parameters still exist.To remedy the problem of poor correlation between 3D feature selection settings and actual target features,a detection method that incorporates point-pillars networks and DETR is proposed.Firstly,pillar coding is employed to offset the influence of redundant point clouds,which improves both computational efficiency and the match of feature extraction.Secondly,the prediction module based on the DETR decoder adopts a multi-headed attention mechanism to establish the correlation mapping between global features and the prediction set,and the prediction results with the strongest correlation are produced in parallel,avoiding the uncertainty caused by manual reliance on the intervention parameters of priori knowledge.Finally,the method is validated on the publicly available dataset,and the mAP is increased by 19.14%and FPS by 3 compared with the original point-pillars network.Also,they both improve substantially when compared to other typical algorithms.