摘要
针对Frustum-PointNets的实例分割网络结构单一且卷积深度较深、易出现特征丢失和过拟合,检测准确率较低的问题,提出了一种改进的Frustum-PointNets网络。该网络首先构建深度残差网络并融入实例分割网络,提高特征提取能力,解决深层网络的退化问题;引入双重注意力网络以增强特征,提高分割效果;运用Log-Cosh Dice Loss解决样本不均衡,加快网络训练;使用Mish激活函数保留特征信息;最后基于Kitti和SUN RGB-D两个数据集进行实验验证本文算法的有效性。实验结果表明,本文算法相对于Frustum-PointNets,在Kitti数据集中,3D框检测精度提高了0.2%~13.0%;鸟瞰图的3D框检测精度提高了0.2%~11.3%。在SUN RGB-D数据集中,本文算法的3D框检测精度提高了0.6%~16.2%,平均检测精度(m AP)提高了4.4%。实验验证,本文算法在室外和室内场景中获得较好的目标检测及分割效果。
For the example of Frostum-PointNets has a single split network structure, deep convolution depth, and is prone to feature loss and overfitting, and the detection accuracy is low. This paper proposes an improved Frostum-PointNets network. The network first constructs a deep residual network and integrates it into the instance segmentation network to improve the feature extraction ability and solve the degradation problem of the deep network.Introduction of dual attention modules to enhance features and improve the effect of segmentation;Log-Cosh Dice Loss is used to solve sample imbalance and speed up network training;Use the Mish activation function to preserve feature information;Finally, based on the two datasets of Kitti and SUN RGB-D, the effectiveness of the proposed algorithm is experimentally verified. Experimental results show that compared with Frostum-PointNets, the accuracy of 3D frame detection is improved by 0.2%-13% in Kitti dataset. The accuracy of 3D frame detection of aerial view is improved by 0.2% to 11.3%. In the SUN RGB-D dataset, the detection accuracy of the 3D frame of the proposed algorithm is improved by 0.6%-16.2%, and the average detection accuracy(m AP) is increased by 4.4%.Experimental results show that the proposed algorithm obtains better object detection and segmentation effects in outdoor and indoor scenarios.
作者
赵瑞
陶兆胜
宫保国
李庆萍
吴浩
ZHAO Rui;TAO Zhao-sheng;GONG Bao-guo;LI Qing-ping;WU Hao(School of Mechanical Engineering,Anhui University of Technology,Anhui Maanshan 243032,China)
出处
《齐齐哈尔大学学报（自然科学版）》
2023年第1期31-41,共11页
Journal of Qiqihar University(Natural Science Edition)
基金
安徽省自然科学基金面上项目(2108085ME166)
安徽高校自然科学研究项目重点项目(KJ2021A0408)。
