融合随机森林模型的单目视觉人体空间定位方法

Human body localization based on monocular vision fused with random forest model

基于低精度视觉设备布置环境,提出了一种融合随机森林模型的单目视觉人体定位方法。首先,采用OpenPose网络模型提取图像中的人体骨架信息,设计人体跟踪与关联算法,保证人体骨架信息的稳定性与正确性,进而通过设计定位特征筛选条件,从骨架信息中选取合适的人体定位特征;其次,通过分析定位特征与骨架关节点之间的关系,构建并训练以人体定位特征点为样本标签属性的随机森林回归模型;然后,融合获取的目标骨架信息以及训练所得的随机森林模型,计算人体空间定位特征参数,并采用视觉相机几何约束模型实现对人体目标位置的定位;最后,搭建单目视觉人体目标定位平台,验证所提方法的可行性及有效性,所提方法采用实时预测等效焦距的方式,其径向定位平均误差为4 cm,在存在遮挡等极端问题时,平均误差也可以控制在10 cm以内。

Propose a human body localization method based on the fusion of monocular vision and random forest model under low-precision hardware condition. Firstly, the OpenPose network model is employed to extract human body skeleton. The target tracking and classification algorithm is studied to ensure the stability and correctness of the human body skeleton. The appropriate human body localization features are selected from the body skeleton by designing the localization feature selection condition. Secondly, by analyzing the relationship between the localization features and the body skeleton joints, a random forest regression model is formulated and trained with the human spatial localization feature points as the sample label attribute. Thirdly, the target body skeleton and the trained random forest model are combined. In this way, the final human body spatial localization characteristic parameters are calculated. The visual camera geometric constraint model is used to locate the human body spatial target. Finally, a space localization platform of monocular vision for human body is established to evaluate the feasibility and effectiveness of the proposed method. The average radial positioning error is about 4 cm by using the method of predicting the equivalent focal length in real time, and the average error can be also controlled within 10 cm when there are extreme problems such as occlusion.