面向边缘GPU设备的快速光流估计算法

Fast optical flow estimation algorithm for edge GPU devices

  • 摘要: 提出了一种适用于边缘GPU设备的光流估计网络,旨在解决稠密光流估计由于巨大计算量而难以在嵌入式系统上部署的问题。首先,针对充分发挥GPU资源的需求,设计了一个高效的特征提取网络,以降低内存访问成本;其次,通过采用扁平形结构的迭代更新模块来估计光流,进一步缩小了模型的尺寸,并提升了GPU带宽的利用率。在不同数据集上的实验结果表明,本文模型具备高效的推理能力和出色的光流估计能力。特别地,与目前先进的轻量级模型相比,所提出的模型在仅使用0.54 Mb参数的情况下,误差减少了12.8%,推理速度提升了22.2%,在嵌入式开发板上展现出了令人满意的性能。

     

    Abstract: An optical flow estimation network suitable for edge GPU devices was proposed, aiming to solve the problem that dense optical flow estimation was difficult to deploy on embedded systems due to huge quantity of computation. Firstly, to fully exploit the GPU resources, an efficient feature extraction network was designed to reduce memory access costs. Secondly, by adopting a flat-shaped iterative update module to estimate the optical flow, the size of the model was further reduced, and the utilization of GPU bandwidth was improved. Experimental results on different datasets show that the proposed model has efficient inference capability and excellent flow estimation performance. In particular, compared with the advanced lightweight models, the proposed model reduces the error by 12.8% with only 0.54 Mb parameters, and improves the inference speed by 22.2%, demonstrating the satisfactory performance on embedded development boards.

     

/

返回文章
返回