一种复杂背景下多尺度多视角的人体检测方法

摘要

Aiming at the problems such as noise interface,complex background and mutual occlusion encountered by human body detec-tion technology under complex background,we designed a multi-scale and multi-view body detection algorithm.According to the shortcomings of traditional object feature extraction method of orientated gradient histogram including high feature-dimension and low detection rate while being occluded,in extraction we employed the extended multi-scale orientation feature and the multi-scale histogram of orientated gradient co-ded by WTA hash separately,and used weak classifier and greedy algorithm to select features so as to obtain the coarse features and fine fea-tures of the image.After that we then used linear shift to synthesise the multi-view samples.The multi-level cascade Adaboost algorithm and support vector machine were used as the classifiers to detect body objects,and the detection accuracy was improved in combination with com-plex background processing and characteristics reinstalling.Experimental results on INRIA public test set showed that the algorithm can make accurate detection on human body objects with multi-view and multi-pose under the conditions of complex background and mutual occlusion. Compared with traditional human body detection algorithm,it has higher detection efficiency and accuracy.%针对复杂背景下的人体检测技术所面临的噪声干扰、背景复杂、相互遮挡等问题，设计一种多尺度多视角人体检测算法。针对传统的梯度方向直方图目标特征提取方法特征维数大、有遮挡时检测率低等缺陷，分别使用扩展多尺度方向特征和经 WTA hash 编码的多尺度梯度方向直方图特征提取，并使用弱分类器和贪婪算法进行特征选择以获得图像的粗特征和精特征。然后使用线性平移合成多视角样本，使用多层级联的 Adaboost 算法和支持向量机作为分类器进行人体目标检测，结合复杂背景处理、特征重装等方法提高检测精度。使用 INRIA 公共测试集的实验结果表明，该算法可精确检测出复杂背景下相互遮挡情况下多视角、多姿态的人体目标，与传统的人体检测算法相比，具有更高的检测效率和检测精度。