基于视觉注意模型化计算的行人目标检测

Pedestrian detection based on modeling computation of visual attention

提出一种用于行人目标检测的视觉注意模型化计算方法。在提取行人目标显著表象特征的基础上,通过多尺度分解后,各尺度图像之间的差减与归一化运算构成显著图;同时,根据肤色高斯似然计算模型,提取行人目标特有的皮肤颜色特征图,再通过分块图像中分类像素点累计计数与阈值化均值滤波相结合的方法精化肤色提取结果以构成导引图;进而提出一种将显著图与导引图通过有偏置的加权整合策略实现目标区域的准确预测。通过Penn-Fudan行人数据库和实拍视频的实验结果表明,所提方法的检测准确率优于现有其他计算模型,且相对传统目标检测算法,能够大幅度减少时间开销,提高检测效率。

A pedestrian detection method is proposed based on the modeling computation oI visual attention. First, a saliency map is generated through multi-scale center-surround computation and normalization, based on extraction of saliency feature of pedestrian object. Meanwhile, a skin color feature map is computed based on the skin color Gaussian model to express the unique character of pedestrian object. A guided map is achieved through refining the skin color features by implementing accumulation of specific pixel for every block and threshold average filtering. Moreover, the objective area prediction is completed by offset weighted of saliency map and guided map. Experimental results on Penn-Fudan pedestrian database and real videos show that the proposed computational model outperforms other existing models in terms of detection precision. Compared with traditional object detection method, the proposed method can save time drastically and improve detection efficiency.