Pedestrian detection method using local feature based on vision attention

Liu Juntao; Liu Wenyu; Wu Caihua; Li Xiongwei; Feng Bin

doi:10.11834/jig.20120311

Views : 0 下载量: 418 CSCD: 0

PDF
Export
Share
Collection
Album

Pedestrian detection method using local feature based on vision attention
Vol. 17, Issue 3, Pages: 370-379(2012)
Published：2012
DOI： 10.11834/jig.20120311
稿件说明：

移动端阅览

DOI：

Liu Juntao, Liu Wenyu, Wu Caihua, Li Xiongwei, Feng Bin. Pedestrian detection method using local feature based on vision attention[J]. Journal of Image and Graphics, 2012, 17(3): 370-379. DOI： 10.11834/jig.20120311.

摘要

在复杂背景下检测行人

具有重要的理论和应用价值。为了适应此类场景中光照的变化和行人姿态的多样性

依据人眼视觉注意原理

提出基于视觉注意的局部特征。该特征具有光照和旋转不变性

并能用于多尺度分析。采用基于特征块的行人表示模型

行人被表示为特征块的集合。每一个特征块用基于视觉注意局部特征的统计直方图和位置关系表示。用聚类的方法得到基于特征块的行人模型。依据每一个特征块在检测窗口中的最大响应训练AdBoost检测分类器

并用困难负样本和可信样本提高检测分类器的性能。用滑动窗口方法在图像和尺度空间中找到检测分类器的局部最大响应

以确定行人位置。实验结果表明

与现有方法相比

本文方法对竖直边缘不敏感

可以处理一定程度的遮挡以及姿态变化。

Abstract

Pedestrian detection in images with complex backgrounds is valuable in theory and applications.To deal with the variability of illumination and pedestrian’s poses

the local feature based on vision attention (LFVA)

derived from a visual saliency mechanism

is proposed in this paper.LFVA is illumination and rotation invariant

and can also be used for multi-scale analysis.A pedestrian model based on feature blocks is proposed

in which the pedestrians are represented by a set of feature blocks.Every feature block is represented by the position and histogram of the LFVA.The pedestrian model is obtained by clustering.The AdBoost detection classifier is trained using the maximal responses of the feature blocks and is improved using hard negative samples and trusted samples.The local maximal of the response of the detection classifier in image and scale space is located as the position of the pedestrian by sliding window search.Compared with the existed methods

the proposed method is less sensitive to vertical edges and can deal with occlusion and pose variation to some extent.