情報処理学会 第81回全国大会講演要旨

Multiple object tracking is a mid-level task in computer vision, it aims to locate multiple objects in the input video, and maintain their identities and trajectories. Existing methods are hard to track objects during occlusion, due to the tracker easily lost occluded object parts. In this research, we try to solve the occlusion issue with a novel separated structure, which is based on deep learning method. We separate the occlusion scene into foreground and occluded parts, and pursuit the former part by an ordinary tracker and the later part by a novel detector, which is trained on occluded images, therefore more robust to occlusion.