【摘要】 目的 现有的高精度双目立体匹配网络因计算开销较高，难以部署到算力受限的边缘计算设备上，极大地限制了双目立体视觉系统的适用场景。针对此问题，提出一种轻量级实时立体匹配网络框架（light-weight binocular stereo matching framework, LBSM）。方法 首先，摒弃了高计算开销的“4D代价体+3D卷积”的主流方案，仅采用2D卷积与轻量化通道注意力机制构建了高效的融合注意力（merged attention, MA）代价聚合模块，在降低计算开销的同时减少了信息丢失，使得代价聚合过程更加高效；其次，提出空间自适应视差传播（adaptive disparity propagation, ADP）策略以替代双线性插值，以极低的代价（仅增加约0.03 M参数、0.36 G“乘法—累加”操作数和1.1 ms推理时间）将端点误差和1像素误差分别降低了21.54%和20.73%，揭示了视差上采样策略在轻量级模型中的重要性。在上述方法的基础上，构建了仅依赖2D卷积、无需任何3D卷积层的轻量级模型框架LBSM，通过简单配置即可生成系列模型。结果 实验结果表明，所提模型在大型数据集Scene Flow上以更低的计算开销取得了更高的准确性（LBSM-L以68.5%的“乘法—累加”操作数取得了明显高于BGNet(bilateral grid learning network)的准确性），而且在真实道路场景数据集KITTI 2012(Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago)和KITTI 2015上分别取得了低至2.41%和2.52%的错误率，显著优于ADCPNet(adaptive disparity candidates prediction network)、P3SNet(parallel pyramid pooling stereo network)等主流轻量级立体匹配模型。此外，在嵌入式AI(artificial intelligence)硬件平台上的部署实测显示，模型处理速度达4.59～20.29帧/s。结论 LBSM在计算开销与准确性之间实现了更好的权衡，能够在低算力边缘计算设备上实时完成双目立体匹配任务，在算力受限的实际场景中具有较大的应用潜力。更多还原

【Abstract】 Objective Computer binocular stereo vision draws inspiration from the fundamental principle of human binocular perception of object distance.It has significant application potential in numerous fields,including intelligent manufacturing,autonomous driving,robotic visual navigation,aerospace,geographic remote sensing,smart healthcare,and virtual/augmented reality,and is being increasingly widely adopted.In recent years,stereo matching methods based on deep learning have leveraged the powerful learning capabilities of deep neural networks to learn implicit rules of stereo matching from a large number of training samples.These kinds of methods integrate the entire stereo matching process,including feature extraction,cost volume construction,cost aggregation,disparity regression,and disparity refinement,into an endto-end deep neural network.By leveraging the powerful learning capabilities of deep neural networks,these methods effectively address challenges posed by various complex factors.Currently,the accuracy of deep learning-based stereo matching networks has significantly surpassed that of traditional methods,leading to significant advancements and a substantial improvement in the accuracy of stereo matching.However,as a pixel-level dense matching task,stereo matching inherently involves high computational complexity.Deep learning-based stereo matching models typically require significant computational resources to achieve good disparity accuracy.Existing state-of-the-art stereo matching networks often demand tens to hundreds of giga multiply-accumulate operations to process a pair of stereo images with a resolution of 540×960.In many real-world application scenarios,owing to constraints such as hardware costs and power consumption,stereo matching models often need to be deployed on low-power edge devices.This issue imposes stringent requirements on the models:They must not only achieve good disparity accuracy but also exhibit extremely low computational overhead.Present advanced stereo matching networks generally have high computational costs,making their deployment expensive and unsuitable for such practical scenarios.To address this issue,this study proposes a lightweight binocular stereo matching framework named LBSM,which is designed for low-power edge computing devices.Method The construction and aggregation of cost volume are two critical steps that significantly influence the overall accuracy and efficiency of the process.Thus,in this study,the popular architecture of "4D cost volume+3D convolution" with high computational overhead is abandoned.First,the 4D cost volume is directly reshaped into a 3D cost volume by fusing the channel and disparity dimensions,and only 2D convolutions are used for cost aggregation.In this way,the entire cost aggregation process only requires the use of 2D convolution,which significantly reduces computational overhead while minimizing information loss.Second,a lightweight channel attention mechanism is introduced for cost aggregation,avoiding the stacking of many convolutional layers and making the cost aggregation process efficient.Third,a two-stage network following the "coarse-to-fine" architecture is adopted,in which a spatially adaptive disparity upsampling strategy replaces bilinear interpolation,enabling adaptive propagation of low-resolution disparity estimates with minimal computational cost.This upsampling strategy not only significantly enhances the accuracy of lightweight models with minimal computational overhead but also enables us to reduce the number of "coarse-to-fine" iterations.Finally,on the basis of the above methods,the proposed LBSM is constructed without any 3D convolutions.Through straightforward configurations,LBSM can yield a series of models that are capable of achieving real-time stereo matching on low-power edge computing devices.Result Ablation and comparative experiments on the large-scale Scene Flow dataset demonstrate that LBSM achieves high disparity accuracy with lower computational overhead.Compared with existing lightweight stereo matching models and even some nonlightweight models,the proposed network framework exhibits clear advantages in accuracy and real-time performance.On the real-world road scene datasets KITTI 2012 and KITTI 2015,the proposed method achieves error rates as low as 2.41% and 2.52%,respectively,significantly outperforming mainstream lightweight stereo matching models such as ADCPNet and P3SNet.Additionally,deployment and testing on embedded AI hardware(i.e.,NVIDIA Jetson TX2) platforms show that LBSM achieves much better accuracy with faster processing speed(4.59—20.29 frames per second).Conclusion Benefiting from the aforementioned strategies,LBSM relies solely on 2D convolutional layers and exhibits very low computational overhead.It can perform binocular stereo matching tasks in real time on low-power edge computing devices,achieving a better trade-off between computational cost and accuracy than existing models and thus endowing it with significant application potential in real-world scenarios with limited computational resources.更多还原