LINEAR DYNAMIC DATA FUSION TECHNIQUES FOR FACE ORIENTATION ESTIMATION IN SMART CAMERA NETWORKS

摘要

Face orientation estimation problems arise in applications of camera networks such as human-computer interface (HCI), and person recognition and tracking. In this paper, we propose and compare two collaborative face orientation estimation techniques in smart camera networks based on fusion of coarse local estimates in a joint estimation model at network level. The techniques employ low-complexity methods for in-node face orientation and angular motion estimation to accommodate computational limitations of smart camera nodes. The local estimates are hence assumed coarse and prone to errors. In the joint refined estimation phase, the problem is modeled as a discrete-time linear dynamical system, and Linear Quadratic Regulation (LQR) and Kalman Filtering (KF) methods are applied. In the LQR-based analysis, the spatiotemporal consistency between cameras is measured by a cost function, which is composed as a weighted quadratic sum of spatial inconsistency, input energy, and in-node estimation error. Minimizing the cost function through LQR provides a robust closed-loop feedback system that successfully estimates the face orientation, angular motion, and relative angular differences to the face between cameras. In the KF-based analysis, the confidence level of each local estimate is used as a weight in the measurement update. This model can be further extended to missing data cases where not all local estimates are collected in the network, hence offering flexibility in communication scheduling between the nodes. The proposed technique does not require camera locations to be known a priori, and hence is applicable to vision networks deployed casually without localization.