摘要： 针对目标检测算法的复合检测模板与变形约束进行研究,在目标外观状态满足高斯分布的前提下,提出一种结合隐支持向量机(LSVM)和混合高斯模型(GMM)的目标检测算法。使用滑动窗算法提取检测目标的梯度方向直方图特征,通过引入二次损失函数,将LSVM在目标检测训练过程中的半凸约束问题转化为凸优化问题,并利用GMM获得目标检测的全局优化结果。实验结果表明,相比双树分枝界限算法和DPM算法,该算法具有更高的目标检测准确率。

关键词: 隐支持向量机, 混合高斯模型, 多目标检测, 变形约束, 半凸优化

Abstract: The composite detection template and deformation constraint problems of object detection are studied in this paper.On the premise that the object appearance meets Gaussian distribution,an object detection algorithm is proposed based on Latent Support Vector Machine(LSVM) and Gaussian Mixture Model(GMM).It utilizes sliding window algorithm to extract Histogram of Oriented Gradient(HOG).The semi-convex constraint of LSVM is transformed into the convex optimization problem by introducing the quadratic loss fuction in the training phase.Then the global object detection optimization results are obtained through GMM.Experimental results show that the proposed algorithm has higher object detection accuracy than Dual Tree Branch and Bound(BB) algorithm and Deformable Part Model(DPM) algorithm.

Key words: Latent Support Vector Machine(LSVM), Gaussian Mixture Model(GMM), multiple object detection, deformation constraint;semi-convex optimization

中图分类号: 

• TN911.73