双目立体匹配是计算机视觉领域一个备受关注的研究问题，在工业检测、自动驾驶、机器人导航等系统中获得了广泛应用。基于深度学习的双目立体匹配技术近期在计算精度方面取得了显著进展，但决定其能否在各类智能终端搭载的实时设计与硬件架构研究相对薄弱。双目立体匹配深度神经网络的计算、数据和存储都相当密集，存储带宽、系统能耗等问题仍是制约该技术实时性能的主要瓶颈。本课题以双目立体匹配的技术需求为背景，以提出兼顾计算精度和硬件效率的电路架构为目标，研究双目立体匹配深度神经网络的算法、架构与电路协同设计问题。重点研究双目立体匹配轻量化网络模型，从系统顶层降低电路负荷；提出动态可重构的卷积计算引擎架构以支持多种卷积操作，提高电路资源的利用率；设计高存储效率的双目立体匹配深度神经网络系统架构，提高系统整体的硬件效率。本课题对双目立体匹配深度神经网络的轻量化模型算法及其高能效硬件架构研究具有重要的学术意义和应用价值。

Binocular stereo matching is one of the most active research topics in computer vision, which has been widely used in a wide variety of application areas, including industrial inspection, unmanned vehicles and robot navigation. Recently deep learning based binocular stereo matching has achieved significant improvement on computational accuracy. However, the real-time processing and hardware architecture design of this technology which determine its ability to deploy in intelligent system receive very little attention. The computation, data and memory of the end-to-end stereo network are quite intensive, thus the memory bandwidth and power consumption are still the major bottlenecks to the real-time performance of this technology. This project aims to exploit algorithm, architecture and circuit co-design to balance the computational accuracy and hardware efficiency of the deep learning based stereo matching. The project firstly focuses on the lightweight network model of stereo matching, which reduces the hardware burden from algorithm level. To improve the utilization of the hardware resource, we then propose a dynamic reconfigurable architecture for convolutional computation engine to support various convolutional operations. Finally, a high memory efficient architecture of stereo network is designed to improve the hardware efficiency of overall system. The research work in the project is of great importance to both academic research and industrial application for the lightweight algorithm and high efficient architecture design of the end-to-end stereo network.