基于边缘设备的快速单目深度估计算法研究  

Research on Fast Monocular Depth Estimation Algorithm Based on Edge Devices

作　　者：王文帅 韩军[1] 邹小燕 倪源松 胡广怡 WANG Wenshuai;HAN Jun;ZOU Xiaoyan;NI Yuansong;HU Guangyi(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China;Zhejiang Huayun Information Technology Company,Hangzhou 310051,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444 [2]浙江华云信息科技有限公司,杭州310051

出　　处：《计算机测量与控制》2025年第4期262-269,共8页Computer Measurement &Control

基　　金：国家自然科学基金项目(62371278,62371279)。

摘　　要：单目深度估算采用单一相机,安装方便,在机器人、无人机领域有广泛的应用;由于单目深度估计算法采用基于编码-解码的复杂的深度神经网络结构会导致边缘设备实时推理效率较低的问题,进而提出了一种可以在边缘设备上实时深度估计的网络架构;该架构采用倒置残差块设计的编码端,采用残差深度可分离卷积与最近邻插值重新设计的解码端,大大减少了模型的参数和计算量,并通过跨层连接将编码网络的特征与解码网络的特征相融合增强深度图中物体的边缘细节信息;实验结果表明,提出的网络架构参数量减少了82%,计算量减少了92%,在KITTI数据集上达到了先进的性能,并且在Jetson TX2上推理速度达到了50 FPS。Monocular depth estimation,employing a single camera with easy installation,is widely applied in the fields of robotics and unmanned aerial vehicles.However,it adopts complex depth neural network structures based on encoder-decoder architectures in monocular depth estimation algorithms,which results in lower real-time inference efficiency on edge devices.Consequently,a network architecture is proposed to enable real-time depth estimation on edge devices.This architecture adopts an encoder designed with inverted residual blocks and a decoder redesigned with residual depth-wise separable convolution and nearest neighbor interpolation,significantly reducing the model's parameters and computational load.Moreover,through cross-layer connections,the features from the encoder and decoder networks are fused to enhance the representation of fine-grained edge details in the depth map.Experimental results show that the proposed network structure has an reduction of 82%in model parameters and a reduction of 92%in computational load,achieving state-of-the-art performance on the KITTI dataset.Notably,the proposed architecture achieves a real-time inference speed of 50 frames per second(FPS)on the Jetson TX2 platform.

关 键 词：深度感知 单目相机 边缘设备 倒置残差 神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]