双边特征聚合与注意力机制点云语义分割

Semantic segmentation of point cloud via bilateral feature aggregation and attention mechanism

机器视觉是环境感知的重要手段之一,是自动驾驶、机器人、工业检测等领域的研究热点,而点云数据的精细分析是其中的一项关键技术。针对大尺度真实场景点云数据分割精度低的问题,提出了一种适用于点云数据语义分割的网络结构。首先,构建了一个双边特征聚合结构,通过分别处理点云的几何信息和语义信息,达到充分利用点云特征信息的目的。其次,使用近邻特征的高维空间相关性计算点与点之间的相互作用,进行局部邻域的上下文信息增强。提出了一种混合池化结构代替最大值池化,减少信息损失,使用横向跨层池化连接来增强特征多样性。最后,引入注意力机制提取全局特征,滤除尺度噪声,增强特征在空间上的表现力。实验结果表明,该方法在大尺度真实场景点云数据集S3DIS上的平均交并比为68.2%,平均准确率为80.7%,比PointNet提高了20.6%和14.5%,客观指标优于已有的代表性方法。

Machine vision is one of the important measure manners for environmental perception.It is a research hotspot in the fields of automatic driving,robot,industrial detection and so on.The fine analysis of point cloud data is one of the key technologies.To solve the problem of low segmentation accuracy of large-scale point cloud data of real scene,a bilateral feature aggregation network architecture for semantic segmentation of the point cloud is proposed.Firstly,a bilateral feature aggregation module is formulated to aggregate local features by processing the geometric information and semantic information of the point cloud.The aim is to make full use of the feature information of the point cloud.Secondly,the high-dimensional spatial correlation of nearest neighbor features is used to calculate the impact between points.The context information of local neighborhood is enhanced.A hybrid-pooling architecture is proposed to replace the max-pooling to reduce the information loss of max-pooling,and the horizontal skip connection pooling is used to enhance feature diversity.Finally,an attention module is introduced to extract global features,which can filter scale noise and enhance the spatial expressiveness of features.Experimental results show that the mean intersection over union of the proposed method is 68.2%,and the mean accuracy is 80.7%.These two values are 20.6%and 14.5%higher than those of the PointNet.The objective indicator is better than the existing representative methods.