一种利用特征选择改进的行人检测模型

An improved human detection model using feature selection

标准HOG模型在行人检测领域中最为经典,相比于标准模型中整齐划一的block,不同尺寸的block可以获得更多的细节信息。首先,在去除上下文背景的32×96尺寸模型基础上设计144个block特征;然后,提出类Fisher比计算block类别区分力;最后,利用NMS方法选出24个block,串接为1 854维的行人检测模型。实验结果表明,该利用特征选择改进的行人检测模型获得了显著的性能提升。

Standard HOG model is the most classic model in the field of human detection.Compared to uniform blocks in the standard model,blocks with different sizes can get more details.Firstly,144 blocks were designed on the basis of the 32×96model which the context of standard model was removed.Secondly,Fisher-like ratio was proposed to calculate blocks＇ discrimination performance.Finally,24 blocks were selected by NMS feature selection method and composed a 1854-dimensional human detection model.The experimental results indicate that the improved human detection model using feature selection achieves significant performance improvements.