| 上村 敬志 | D, 中間発表 | サイバネティクス・リアリティ工学 | 清川 清, | 向川 康博, | 内山 英昭, | Perusquia Hernandez Monica, | 平尾 悠太朗 |
|
title: title : 3D Gaze Direction Estimation from Surveillance Camera Videos Based on Full-Body Posture and Temporal Rotation Modeling
abstract:Surveillance cameras in public and commercial spaces provide information about human position, posture, and behavior. Gaze direction is a useful cue for estimating attention. However, 3D gaze estimation in surveillance videos is difficult because people are often observed from elevated and distant viewpoints, at low resolution, from non-frontal directions, or under occlusion. Methods relying on high-resolution face or eye images cannot be directly applied. This dissertation studies 3D gaze direction estimation from surveillance videos using full-body posture and temporal rotation modeling. The research has three stages. First, a geometry-based method estimates 3D posture and approximates gaze from head keypoints, showing feasibility without eye appearance but also limitations due to pose errors and the head-gaze assumption. Second, an end-to-end neural network estimates gaze from a single frame using full-body and 3D pose information. Third, a multi-frame framework estimates reference gaze, models gaze changes as SO(3) rotations, represents uncertainty with Matrix-Fisher distribution, and aggregates the sequence with LSTM. It achieves the lowest error among compared methods. language of the presentation: Japanese | |||||||