基于深度学习多目标跟踪.pdfVIP

Object Tracking by Multi Instance Depth Learning and Online Learning of Non-linear Motion Patterns Yuxiao Zheng School of Electronic Information, Northwestern Polytechnical University Abstract Target tracking technology is an important part of computer vision research field, in order to solve the drift of the target or loss of tracking problems such as tracking under complex scene, this paper proposes the combination of the appearance information and motion information of moving objects. A multi instance depth learning algorithm is used to construct a stable apparent model and a online learning of Non-linear motion patterns algorithm is used to construct motion map.We evaluate our approach on public data set, and show significant improvements compared with state-of-art methods. Keyword: Target tracking, A multi instance depth learning, online learning of Non-linear motion patterns. 1. Introduction Target tracking is an important part of computer vision research field, n real scenes is important for manyapplications, such as surveillance, robotics, and human-computer interactions. and so on, but still faces many challenges. This problem becomes difficult incrowded scenes with frequent occlusions and interactionsamong the targets, as well as the target shape and illumination changes, these factors will lead to the target drift and even track loss. Deep learning is a new research field in machine learning, The method of deep study to construct the depth of the network to extract features is the research direction of the target and behavior recognition. Many studies have indicated that deep learning can learn from a large number of data, which has a great help for the detection and