面向“最后一公里”的智能驾驶是人工智能技术的一个重要落地应用方向。而基于视觉的泊车位检测系统是这其中的关键一环，它是短程自主泊车、面向停车场环境的视觉SLAM等高层系统的核心模块。本课题旨在解决基于视觉的泊车位检测技术中的关键问题，将围绕环视相机系统的离线标定、环视相机系统的在线动态补偿、环视图像中的泊车位检测、大规模泊车位图像数据集的构建四个方面展开探索。课题提出新的高效高精度环视相机系统标定流程以及多相机位姿联合优化算法，提出以最小化光度误差为目标的相机位姿修正与鸟瞰视图光度配准联合优化技术框架，提出面向环视泊车位检测问题的多属性深度回归神经网络模型，探索大规模仿真泊车位环视图像数据集的构建机制，提出引入语义相似度正则化约束的生成对抗网络模型来有效提升仿真泊车位图像样本的真实性。本课题的开展将有效促进计算机视觉技术和人工智能技术在智能驾驶领域的应用和发展，促使基础研究成果走向应用。

Intelligent driving for the “last kilometer” is an important application direction of artificial intelligence technology. The vision-based parking-slot detection system is one of its key parts. The vision-based parking-slot detection system is the core module of some high-level systems, such as short-range autonomous parking and visual SLAM for parking environment. This project aims to solve the key problems in vision-based parking-slot detection technology. It will focus on four aspects: offline calibration of the surround-view camera system, online dynamic compensation of the surround-view camera system, parking-slot detection based on the surround-view image, and the construction of the large-scale dataset of parking-slot images. We propose: 1) a new high-efficiency and high-precision pipeline for calibrating a surround-view camera system and an algorithm for jointly optimizing poses of multi-cameras; 2) a joint optimization framework for correcting cameras’ poses and registering bird’s-eye-view images in photometric domain by minimizing the photometric error; and 3) a multi-attribute regression deep neural network model for detecting parking-slots in surround-view images. In addition, this project will explore the mechanism for constructing a large-scale dataset of synthetic parking-slot surround-view images. We propose a GAN model with semantic similarity regularization to effectively improve the reality of synthetic parking-slot images. The development of this project will effectively promote the application and development of computer vision technology and artificial intelligence technology in the field of intelligent driving, and promote the application of outputs from basic research.