【作者】 张晨；

【导师】 陈驰； 杨必胜；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2022， 硕士

【摘要】 场景深度信息的感知和获取是近几年计算机视觉领域的研究热点,有效且准确的深度信息可以提升相关应用的质量效果,例如自动驾驶,SLAM(同步定位与建图),增强现实等。但是现有的深度感知方式,包括双目视觉、结构光相机和激光雷达传感器都有一定的缺陷和局限性。为了克服硬件设备存在的技术缺陷,通过算法补全深度图像中缺失像素深度信息的深度补全方法应运而生。随着卷积神经网络在深度补全任务上的应用,补全精度和算法性能都有明显的提升。但目前深度补全方法都存在着一定的问题,比如无法补全过于稀疏的激光点云区域、补全深度边界模糊以及前后景深度混合等。针对以上问题,本文设计了一种多重注意力非局部特征融合的序列影像-激光点云深度补全方法。具体研究内容如下:(1)针对传统卷积神经网络对有效特征提取不足,和噪音明显的问题,设计了一种注意力机制融合的自校准卷积深度补全网络。在编码阶段引入自校准卷积模块和多重注意力机制模块,通过多级下采样,从序列影像和激光点云中获得丰富的场景特征和深度特征。解码阶段利用跳连接,在多尺度阶段融合来自编码阶段的特征,最大程度地获取有效特征,减少噪声。(2)针对深度补全结果存在深度边界上混合深度的问题,使用了一种非局部特征空间传播卷积神经网络补全深度优化网络。根据学习到的非局部邻域和相关性,在初始预测深度信息的置信度指导下,通过非局部特征空间传播过程迭代地优化初始预测的深度补全结果。该网络能够有效地排除不相关的邻域,并将传播集中在相关邻域之间。(3)本文在公开的深度补全数据集上进行实验和分析,实验表明,本文方法能够获得有效且准确的稠密深度图。与现有经典算法对比,在定性结果表现上,本文方法深度边界清晰,有效改善了混合深度的问题,在定量结果上,本文方法补全深度信息准确,有效降低了误差。更多还原

【Abstract】 The perception and acquisition of scene depth information is a hot research topic in computer vision in recent years.Effective and accurate depth information can enhance the quality in related applications,such as Autonomous Driving,SLAM(Simultaneous Localization and Mapping),Augmented Reality,etc.However,existing depth perception means,including binocular vision,structured light cameras and Li DAR sensors,have inherent defects and limitations.To overcome the technical shortcomings of hardware devices,depth completion methods that complement the missing depth information for pixels in depth images through algorithms have emerged.With the utilization of convolutional neural networks in the depth completion task,the accuracy and performance of the completion have been significantly improved.However,there are still some problems in existing depth completion methods,such as the impossible completion in the regions with too sparse laser point cloud,blurred depth-completed boundaries and mixed depth of the foreground and background.To address these problems,this paper designs a depth completion network with multiple attention and non-local feature fusion on image sequences and laser point cloud.The details of the research are as follows:(1)To tackle the problems of insufficient effective features,and obvious noise from traditional convolutional neural network,a self-calibration convolutional depth completion network fused with attention mechanisms is designed.The self-calibration convolutional module and the multiple attention module are introduced in the encoding stage to obtain sufficient scene and depth features from image sequences and laser point clouds by multi-level down sampling.In the decoding stage,the features from the encoding stage are fused on the multi-scale by skip connection obtain effective features and reduce noise as much as possible.(2)To address the mixed depth of depth boundary in the depth completion results,a non-local feature-space propagation convolutional neural network is proposed to complement the depth optimization network.Based on the learned non-local neighborhoods and the corresponding affinity information,the initially predicted depthcompletion results are iteratively optimized by a non-local feature-space propagation process guided by the confidence of initial depth.The network can effectively exclude the irrelevant neighborhoods and concentrate the propagation among the related neighborhoods.(3)In this paper,experiments and analyses are conducted on publicly available depth completion datasets.And the Experiments indicate that the proposed method can obtain effective and accurate dense depth maps.Compared with existing classical algorithms,the depth boundary predicted by the proposed method is clear in terms of qualitative results,which effectively alleviates the problem of mixed depth.In quantitative results,the completed depth information of the proposed method is accurate,which effectively reduces the error.更多还原