基于位置度量关系的自动驾驶3D目标检测

3D Object Detection for Autonomous Driving Based on Position Metric Relationships

摘要:

针对现有的基于激光雷达的3D目标检测网络中的损失函数存在优化难度较大，计算成本高，检测精度低的问题，提出一种改进的激光雷达3D目标检测网络。首先，使用全新的交并比损失函数来更好的度量预测框和真实框的位置关系；随后，优化现有网络的特征编码方式以获取更丰富的点云特征；最后，对锚框正负样本阈值设定进行改进实现更合理的正负样本分配。针对自动驾驶领域，利用KITTI数据集进行相关的实验验证。实验结果显示，改进网络相对于基准网络在KITTI数据集行人和骑行者等小目标分别检测精度提升6.57%，4.60%，平均检测精度提升5.81%。

Abstract:

An enhanced LiDAR 3D object detection network is proposed to address the existing Lidar-based 3D object detection networks ,which suffer from challenges in terms of optimization complexity, high computational demands, and low detection accuracy. First, a new intersection over union loss function is used to measure the positional relationship between predictions and ground truths. Subsequently, optimize the feature encoding method of existing networks to obtain richer point cloud features. Finally, the threshold setting of the anchors is improved to achieve a more reasonable allocation of positive and negative samples. For the field of autonomous driving, the KITTI dataset is utilized for relevant experimental validation. The experimental results show that the enhanced network improves the detection accuracy of small targets such as pedestrians and cyclists by 6.57% and 4.60%, respectively, and the average detection accuracy is improved by 5.81%, relative to the baseline network in the KITTI dataset.