基于可变形部件模型HOG特征的人形目标检测.pdfVIP

Vol. 37 No.3 第 37 卷第 3 期 应用光学 2016 年 5 月 Journal of Applied Optics May 2016 文章编号:1002-2082(2016)03-0380-05 基于可变形部件模型 HOG 特征的人形目标检测 宋 王晋，王世峰 (长春理工大学光电工程学院，吉林长春 130022) 摘 要:使用单幅图像进行特定目标的检测是机器视觉领域的重要任务之一。利用机器学习的 方法，使用 LSVM 分类器进行人形目标的检测。该方法提取图像的 HOG(梯度方向直方图〉特 征和其对应的可变形部件来描述目标的外形特征，能够较好地解决目标由于运动而产生外形变 化的问题。对常见公共区域场景进行数据采集并随机抽取了 200 张图像，使用所述方法对其中 共 1 100 个人形目标进行检测，正确率识别率为 78.3% 。结果表明该方法具有一定的可行性和 稳定性，能够较好检测出单幅图像中的人形目标并加以标泣。但对于某种程度有所遮蔽的人形 目标则会产生汤检的现象。 关键询:可变形部件模型;梯度方向直方图;隐支持向量机;目标检测 中图分类号:TN957.52 文献标志码:A doi : 10.5768/ 1A0201637. 0302003 Human-kind shape object detection using deformable parts model with HOG features Song Jin , Wang Shifeng (Sch∞1 of Opto-electronic Engineering , Changchun University of Science and Technology. Changchun 130022. China) Abstract: It is one of the important tasks in machine vision field to detect specific object using single image. A machine learning approach using latent support vector machine(LSVM) classi- fier was presented for detecting human- kind shape object. The histogram of oriented gradients (HOG) features were extracted and the corresponding deformable parts were formed to de- scribe the appearance of the object. The problem that the object s appearance deformed when it was in motion. was solved. 200 images captured in typical public areas were randomly select- ed and used to perform this method. 1 100 human-kind shape objects from them were tested , and th e correct recognition rate of 78. 3 % was achieved