CAJ | 학술논문

针对单视角下信息量不足以及多视角不同视角间信息关联困难的问题，提出了基于结构化约束的多视角人体检测方法。首先通过基于块的人体检测模型获取人体局部块信息；然后采用空间仿射变换将不同视角下重叠区域通过变换矩阵的映射关系关联起来；最后针对仿射变换后的区域因遮挡或者存在多目标导致多视角目标关联困难的问题，利用人体局部显著块间的结构化约束为多视角目标匹配构造最大后验概率模型，通过最优求解获取多视角目标匹配结果。实验结果表明，该方法能够利用多视角信息来有效弥补单视角下人体检测中出现的遮挡问题，显著提高了人体检测效果。
침대단시각하신식량불족이급다시각불동시각간신식관련곤난적문제，제출료기우결구화약속적다시각인체검측방법。수선통과기우괴적인체검측모형획취인체국부괴신식；연후채용공간방사변환장불동시각하중첩구역통과변환구진적영사관계관련기래；최후침대방사변환후적구역인차당혹자존재다목표도치다시각목표관련곤난적문제，이용인체국부현저괴간적결구화약속위다시각목표필배구조최대후험개솔모형，통과최우구해획취다시각목표필배결과。실험결과표명，해방법능구이용다시각신식래유효미보단시각하인체검측중출현적차당문제，현저제고료인체검측효과。
To solve the problems of lack of information in single-view and the difficulty in information correspon-dence in different views,a multi-view body structure-constrainted human detection method was proposed. First, part-based human detection model is implemented to obtain the information on human body part. Then leverage spatial affine transform to correlate the overlapping regions in different views. Finally,to overcome the challenge of object corresponding in multi-view environment caused by partial occlusion and multiple target existence in neighborhood, the model of maximum a posterior(MAP)isdeveloped for multi-view object matching by taking advantage of the body structure constraints. The multi-view object matching result can be achieved by optimizing the objective function of the modal.The experimental results show that the proposed method can improve human detection by efficiently using multi-view cues to avoid partial occlusion in single-view.