深度学习的街景行道树自动识别提取研究

针对城市行道树调查中,街景影像背景环境复杂多变、行道树个体差异大,依靠目视判读费时费力的问题,该文基于车载移动测量系统采集的全景影像数据,利用深度学习算法,在快速区域卷积神经网络的目标检测方法基础上,建立适用于街景行道树检测的深度神经网络模型。模型采用基于共有显著性区域及冗余策略的行道树多示例目标候选区域选择方法,使用车载图像的几何约束进一步筛选合适的候选区域,从而实现行道树目标候选区域的统一选择,提升行道树目标的检测效果。实验结果表明,该文提出的方法能够实现多种行道树的准确自动识别与提取,进而大大降低行道树绿化调查的成本。

Research on automatic recognition and extraction of street trees based on deep learning

In the survey of urban street trees,the background environment of street scene images is complex and changeable,the street trees vary greatly,and the visual interpretation is time-consuming and laborious.Based on the panoramic image data collected by the vehicle-mounted mobile measurement system,this paper used deep learning algorithms to convolve in fast areas.Based on the neural network target detection method,a deep neural network model suitable for street view street tree detection was established.A street tree multi-example target candidate region selection method based on shared saliency regions and redundancy strategies,and the geometric constraints of the vehicle image to further filter suitable candidate regions were adapted in this paper,so as to realize the unified selection of street tree target candidate regions and improve the detection of street tree targets effect.The experimental results showed that the method proposed in this paper could realize the accurate and automatic identification and extraction of various street trees,thereby greatly reducing the cost of street tree greening investigation.