基于视觉语义理解与生成的复杂场景行人感知关键技术研究

With the increasing demand of the country and society for large-scale intelligent pedestrian perception analysis in complex scenes, the visual semantic understanding based pedestrian perception has become a hot yet difficult topic in the research community. It has important theoretical value and research significance. In this project, we will implement a comprehensive study for complex scenarios of pedestrian perception, explore the visual semantic understanding and generating based person image and video data augmentation. We will propose new weakly supervised image semantic segmentation model and person image generating model with limited data. Further, we will develop more suitable solutions for individual identification under complex scene. Towards different pedestrian perception tasks, i.e., including person recognition, attribute analysis, tracking, and action understanding, we will design a systematic solution and optimizing models for joint learning. In this way, we will achieve accurate and reliable pedestrian perception system under complex scenes. This project is expected to improve the effectiveness and efficiency of the pedestrian perception application under complex scenarios. It will promote breakthroughs in the human-centered vision research and the organic combination of popular computer vision based technologies and the important national and social needs.

随着国家和社会对于复杂场景下大范围智能行人感知分析技术日益增长的需求，基于视觉语义理解的行人感知已经成为当前学术界的热点及难点问题，具有重要的理论价值与研究意义。本项目将全面研究面向复杂场景下的行人感知分析关键技术，探索基于视觉语义理解与生成模型的行人图像及视频数据增强理论与方法，提出数据受限情况下的弱监督图像语义分割模型及行人图像生成模型，开发适用于复杂场景下的个体身份识别技术，针对不同的行人感知任务（包括行人身份识别、属性分析、轨迹跟踪、行为理解等）提出系统解决思路，设计并优化不同层次之间模型的联合学习框架，实现在复杂场景下准确可靠的行人感知系统。本项目有望实现大范围复杂场景中的行人感知应用的性能及效率提升，推动以人为中心的视觉研究取得突破，促进计算机视觉热门技术与国家社会重大需求的有机结合。

内容获取失败，请点击重试