基于图像变换的递归式行人错检校验算法

针对现有HOG特征行人检测器容易受到复杂背景环境的干扰而降低检测效率的问题,提出一种基于图像的递归式行人错检校验算法。在保持行人检测器原有结构不变的基础上,对行人检测器的检测结果进行图像变换,变换结果作为新的输入图像,将原有检测器构成一个递归式错检校验处理结构。采用尺寸归一化、均衡和锐化图像变换方法,对行人检测器检测结果不断迭代校验,以达到降低错检率的目的。同时引入模糊决策判决终止条件,实现合理的迭代退出。经INRIA数据集和实测图片的仿真分析表明,该方法能够有效去除行人检测器输出的错误检测结果,正确检测率在原有基础上提高7.6%,在复杂背景条件下仍有效。

Recursive False Positive Check Algorithm for Pedestrian Detection based on Image Transformation

Due to the interference of complex environmental background, the efficiency of pedestrian detectors using HOG features has declined. To solve this problem, we propose a recursive false positive check algorithm for pedestrian detection based on image transformation. This algorithm makes the pedestrian detector become a recursive false positive check structure, through maintaining the construction of original pedestrian detector and utilizing the detection result as input image of pedestrian detector after image transformation. In order to reduce the false detection rate, image transformation methods such as size normalization, equalization and sharpening are adopted to check the results of pedestrian detector recursively. And fuzzy decision termination condition is introduced to achieve reasonable exit from iteration. The simulation results on INRIA and other collected datasets show that the method can remove the false detection results of the pedestrian detector and the correct detection rate is increased by 7.6% compared with the original detection, which demonstrates the effectiveness of this method in the complex background conditions.