分部加权的行道树检测网络构建

The construction of road tree detection network based on segment weighting

基于深度学习目标检测框架,提出了一种端到端的训练网络,用于行道树的自动检测。由于行道树之间的遮挡问题,现有的通用物体检测框架无法直接应用于此任务,为此本文提出了一种树形分部加权模块,以减少严重遮挡造成的错误检测。然后对提出的神经网络进行训练和评估。结果显示,本文所建立的分部加权树木检测网络能够在遮挡条件下,有效地检测出街景图像中的行道树,该方法在各种条件下均具有较高的精度和良好的稳健性。

In this paper,an end-to-end training network is proposed based on the framework of deep learning forobject detection,which can be used for automatic street treesdetection. Because of the occlusion problem between the roadway trees,the existing general object detection framework cannot be applied to this task directly. In this paper,a tree-shaped partial weighting module is proposed to reduce the error detection caused by severe occlusion. Then,the proposed neural network is trained and evaluated. The results show that the partitioned weighted tree detection network established in this paper can effectively detect the street trees in street scenes under occlusion conditions. The experimental results further show that the method gets high accuracy and good robustness under various conditions.