基于MRF区块聚类模型的人体检测方法

Human detection method based on MRF-blob-clusters model

为解决现有人体检测方法对遮挡、姿态变化人体漏检率高的问题,提出一种鲁棒的人体检测方法。将人体划分为7个区块,针对不同区块提取方向梯度直方图特征,采用支持向量机方法进行训练,得到各个区块的检测器,用于描述相应的区块;将相同类型的区块组成一个聚类,在7个区块聚类上构建马尔科夫随机场区块聚类模型;采用变分法进行推理,求取模型最优解,检测人体目标。实验结果表明,与目前主流的人体检测方法相比,该方法的平均漏检率低,对人体遮挡和姿态变化的鲁棒性强。

For solving the problem that current human detection methods have high miss rate while detecting human with occlu- sion or multi-poses, a robust human detection method was proposed. The human body was divided into seven blobs, histogram of oriented gradient features were extracted from them, support vector machine method was used to train the features, and every blob＇s detector was obtained for describing the corresponding blob. Blobs with same type were combined into a cluster, MRF （Markov random field）-blob clusters model on seven blob clusters was built. The optimal solution was calculated and human was detected using variational calculus. Experimental results show that, comparing to the state-of-the-art methods, this method has lowest logaverage miss rate and strong robustness for occlusion and pose-variation of human.