基于改进YOLO v3的相似外部特征人员检测算法

针对目前异常行为检测中相似特征人员检测方法稀缺、人员和特征检测准确度低、特征量较少,而相似外部特征往往意味着团队行动及潜在异常行为等问题,提出一种基于深度学习的相似外部特征人员检测算法.首先采用加入Fast Guided Filter的暗通道去雾算法对INRIA数据库图像进行前期处理,得到质量更佳的训练样本;然后用得到的样本对改进的YOLO v3进行训练;最后将提取出来的行人进行颜色特征和几种纹理特征提取,组合之后用ELM进行分类.仿真结果表明:加入Fast Guided Filter的暗通道去雾算法明显优于单纯的暗通道去雾算法,保留了更多的边缘和纹理特征,在雾天和强曝光下效果尤为明显.相比HOG+SVM方法,该算法对人员检测的误检率和漏检率都大大降低,且具有较好的实时性.最后ELM分类的准确性能够达到96. 104%.

Similar External Features Personnel Detection Algorithm Based on Improved YOLO v3

Similar external characteristics often mean problems such as team action and potential abnormal behavior. To solve the problems of fewer detection methods for similar features,low accuracy and fewer features in abnormal behavior detection,an algorithm for detecting similar external features based on in-depth learning is proposed. And similar external features often mean team action and potential abnormal behavior. Firstly,the dark channel defogging algorithm with Fast Guided Filter is used to pretreat the image of the INRIA database,and the training samples with better quality are obtained. Then,an improved detection algorithm of YOLO v3 is trained with the obtained samples. Finally,pedestrians will be extracted from the extraction of color features and texture features of several combinations,after using ELM classification. The simulation results show that the dark channel defogging algorithm with Fast Guided Filter is superior to the simple dark channel defogging algorithm,retaining more edge features and texture features,especially in foggy weather and strong exposure environment.Compared with the HOG + SVM method,the false and missed detection rate of pedestrian detection in this algorithm are greatly reduced and have better real-time performance. Finally,the accuracy of ELM classification can reach 96. 104%.