Intelligent Model-Based Tracking of Natural Gait Motion in a Network of Depth Sensors

点击翻译按钮获取中文摘要

Tracking human body by observing its motion patterns in a natural living environment is a fundamental research area with many potential applications. In the past few years, advancement and widespread availability of commodity depth sensors, integrated with cameras have been offering a non-wearable alternative. This research concerns with some aspects of tracking motion of people in a network of sensors located at their natural living environment. Most existing approaches for processing the sensed depth information are incorporating complete or partial 3D depth data. Fundamental to our approach is the intelligent utilization of a collection 2D scan planes of the 3D data as a part of motion and scene representation. As a part of this investigation, it is proposed to investigate methods for determining initial changes in the sensed information due to movements. This detection can be used to assign and schedule sensors in the network and plan paths for mobile sensors. Each of the 1D depth profile in the associated scan planes can be further analyzed using Fourier analysis for defining components of feature vectors for further classification. It is proposed to explore the development of a 2D scan plane planner based on an intelligent particle filter tracking framework. As a part of this estimator, models for the gait movement patterns and postures of the subject are to be incorporated. The model-based estimator compares the sensed feature vectors with stored features associated with a number measured scans and generated key graphical postures and kinematic chain models of various limbs. The kinematic chain model is constructed using hierarchical description of particle filter starting from the kinematics of the torso to more detailed kinematic model of the connected limbs. In this work, we propose to explore the incorporation of visual information of hands and head based on skin color. This additional information will assist in further localization of hand extremities within the depth data for cases when the hand is in contact with the body (torso) or with various other limbs or body parts. We propose to approximate the head orientation during the gait pattern or various activities based on the depth data and face placement. For the cases where body occlusion is inevitable from sensor due to the presence of furniture (such as during the sitting down and raising from a chair), it is proposed to incorporate depth information from 2D scans of the furniture for defining its various feature vectors. We will compare this information with models of the stored features of furniture associated with the living environment to determine both types, position, and orientation of the furniture in relation to the subject.