基于后验多重稀疏字典的多姿态行人检测

Multi-pose Pedestrian Detection Based on Posterior Multiple Sparse Dictionaries

为了有效检测行人,提出一种基于后验多重稀疏字典的多姿态行人检测方法。根据多个不同稀疏度字典对图像稀疏编码,构成分布直方图作为行人图像多重稀疏字典特征;统计全部正样本特征的共性信息,对单个行人样本特征加权,获得表征行人的后验多重稀疏字典特征;根据视角和行人姿态的不同利用聚类方法划分子类,且训练等权重加和方式的多姿态分类器。在多个数据集上的测试表明,所提方法的特征维数低,描述能力优,有效提高了行人检测精度。

In order to detect pedestrians effectively, a multi-pose pedestrian detection method based on posterior multiple sparse dictionaries was proposed. Through pre-learning multiple different sparse dictionaries, and sparse coding the image, statistics for each dictionary corresponds to sparse coding histogram as the pedestrian image feature descriptor. The common information of multiple sparse dictionary features of all positive samples was obtained, and the feature of a single pedestrian sample was weighted, and the features of a posteriori multiple sparse dictionary could be obtained. Then pedestrians of different poses and views were divided into subclasses with clustering algorithm. A classifier was trained for each subclass. A multi-pose-view ensemble classifier was trained to combine the output values of different subclass classifiers with an equally weighted sum rule. Experimental results on different datasets suggest that the proposed posterior feature is more than the classical sparse dictionary and other typical features. Compared with the existing methods, by combining the posterior feature and the multi-pose-view ensemble classifier, the proposed method improves the detection accuracy effectively.