深度图像的目标潜在区域提取算法  

Object Proposal Algorithm for the Depth Image

作　　者：方智文[1,2] 曹治国[1] 肖阳[1] 

机构地区：[1]华中科技大学自动化学院多谱信息技术国家级重点实验室,湖北武汉430074 [2]湖南人文科技学院能源与机电工程系,湖南娄底417000

出　　处：《信号处理》2016年第2期193-202,共10页Journal of Signal Processing

基　　金：国家高技术研究发展计划(863计划)(2015AA015904);中国博士后科学基金资助项目(2014M562028);湖南省教育厅资助科研项目(14C0599);中央高校基本科研业务费资助HUST(2014QNRC035和2015QN036)

摘　　要：目标检测和识别算法通常使用复杂特征以多尺度滑动窗的方式进行分析,运算效率往往非常低。因此,目标性被引入进行目标潜在区域的快速预判断,减少复杂特征需要分析的窗数,从而达到加速算法效率的目的。针对逐步普及的Kinect深度像机,该文提出了一种基于深度图像的目标性分析算法,以提升深度图像的目标检测识别算法的效率。首先基于深度图像的法向量,提出能够有效描述深度图像边缘信息的边缘检测方法,然后通过支撑向量机学习目标性的分类器,以得分的形式给出候选区域中存在目标的概率,最后基于人眼的视觉机理对不同尺度的目标进行加权。通过深度图公共数据库(UW和NYU)的实验对比,该算法给定1000个候选区域时分别达到94.1%和92.9%的召回率,保证了准确率的同时大大减少了区域数量,能有效的提升目标检测识别算法的效率。Based on the multi-scale analysis,the complex features were calculated with the sliding windows in many target detection and recognition algorithms. However the efficiency was low. Aiming to promote the efficiency,the objectness was introduced to pre-analyze the potential location of objects. While the KINECT was widespread,an objectness method for the depth image was proposed to leverage the efficiency of other algorithms in the depth map. Firstly,the edge information was extracted from the normal vector of the depth images. Secondly,SVM was used to classify the object according to the score of the objectness. Finally,the different weights were learned for the different scales based on the visual mechanism. The comparative experiment results in the public database（ UW and NYU） demonstrated that the recall rate of our method achieved 94. 1% and 92. 9% with 1000 proposals respectively. It can leverage the efficiency of the target detection and recognition because of a few candidates with high detection rate.

关 键 词：目标检测 深度图像 边缘检测 支撑向量机 目标性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]