Artificial Intelligence for Visual Processing of Humans in Images and Videos

点击翻译按钮获取中文摘要

Human detection plays a central role in video surveillance, video monitoring, image-based entertainment, etc. Surveillance systems with the ability to understand human activities will be able to recognize activities and detect suspicious actions to prevent undesirable incidents from happening. Likewise, game and movie industries, as well as sports, will benefit from automatically analyzing human activities in videos and segmenting them precisely in images. Moreover, it is vital for decision-making systems such as autonomous cars to understand the intentions of surrounding peoples in real-time. In the context of this project, we aim to combine traditional computer vision technology with modern deep learning-based techniques to push the boundaries in processing human in monocular images and videos. More specifically, we aim to fuse the human pose information with contextual information and model their evolution over time to detect human activities and intentions in untrimmed video. Concretely, there are two challenges to achieve the above goal. First, developing a fast and accurate technique for fusing posture information with contextual information and, second, devising a method for estimation human posture regardless of the complexity of the posture and the context depicted in the image/video.