摘要
针对传统运动方向异常检测方法需要人工参与、智能化较低等缺点,提出了一种新的基于视觉注意机制的运动方向异常检测方法。该方法通过合成源图像的空间导数图与时间导数图获得运动边缘,求取沿空间各方位分布的运动方向特征图。在此基础上,利用所提出的基于运动区域面积的归一化方法对得到的每一幅特征图按照显著性高低分别赋予不同的权重,完成不同运动方向特征的相互竞争。最后融合经过归一化处理的特征图以得到最终的运动显著图。该显著图中运动方向具有显著性的物体得到有效突出,达到了存在多个运动物体的情况时,运动方向上具有显著性的物体能够更加有效、智能地检测出来的目的。
Aiming at the disadvantage of traditional anomaly detection to motion direction such as requirement of human involvement and low intelligence, a novel detection method based on visual attention was proposed. In this method, feature maps of motion direction distributed along every orientation in space were firstly obtained by integrating the spatial and temporal derivatives of source image. On this basis, the new proposed normalization method based on area of motion field was used to endow different weight to all feature maps according to their own saliency, in order to achieve the competition among different motion direction features. The final motion saliency map could be acquired by merging all normalized feature maps. The object with obvious saliency in motion direction was focused in this saliency map, so the destination that detected more effectively and intelligently from multiple object with motion direction saliency could be moving objects.
出处
《红外与激光工程》
EI
CSCD
北大核心
2012年第5期1379-1383,共5页
Infrared and Laser Engineering
基金
教育部新世纪优秀人才支持计划(NCET-08-36)
长江学者和创新团队发展计划(IRT0705)
关键词
视觉注意
运动方向
异常检测
显著性
visual attention
motion direction
anomaly detection
saliency
