Advanced methods for depth-based human pose estimation and motion analysis: application to vital signs monitoring in the intensive care unit

点击翻译按钮获取中文摘要

Despite the mature and affordable solutions offered today in commoditized range sensors such as Microsoft's Kinect with its skeletal tracking algorithm, several challenges remain regarding depth-based human pose and motion analysis. Existing algorithms rely mostly on supervised learning with training data obtained in almost ideal conditions. Considerations with regards to range sensing technology, to subject's clothing and to visual occlusions are barely taken into account in the training. Another question remains regarding the performance of these methods when the subject is sitting or even lying down. Moreover, historically, the primary focus in human motion analysis has been on estimating and/or tracking human body joints over time. Despite its popularity, joint-based representation does not offer enough resolution to analyze fine and local motion such as human tremors, respiratory motion and nodding for example. There is a need in moving beyond the stick-figure view of the human body toward recovering richer descriptions of shape and motion. Building on my previous work, the long term objective of this research program is to propose novel dense representations of human pose and motion from depth images that allow a robust analysis of both fine-local and ample-global motion in real-time and in unconstrained environment. To achieve this, we will develop and validate novel computational tools and methods that address the challenges related to range sensor variability, to posture variability, to fine motion description and to severe occlusions. These tools will then be applied in the context of patients monitoring in the pediatric intensive care unit (PICU) for real-time detection of signs of vital distress. The program is composed of four specific objectives: (SO1) to enhance the robustness of depth-based human pose estimation to sensors specific artifacts and to bed-ridden postures, (SO2) to propose a new semantic dense motion descriptor from a sequence of depth images while taking into account the strategies developed in SO1, (SO3) to propose new strategies for human pose estimation and motion tracking in the presence of sever occlusions, (SO4) to detect abnormal head and limb movements in PICU patients by applying the tools developed in SO2 and SO3. These objectives will train 2 PhD, 3 master and 5 undergraduate students. By tackling important technical challenges, the proposed program will lead to major advances in human pose estimation and motion analysis. The computational tools will benefit a wide variety of applications such as robotics, surveillance, gaming and advanced manufacturing. In-bed motion analysis has received little interest so far although it brings innovative value into the market of commoditized child monitoring systems. Finally, this research program will have a direct impact on the quality of care in the PICU by preventing management delays, improving medical staff's efficiency and thus patients' outcome.