Research on street tree extraction method based on improved RandLA-Net

In view of the challenges faced by street tree extraction in the current urban road scene, such as the problem of accurate segmentation caused by insufficient automation and the overlapping of ground object point clouds, this study is committed to improving the automatic recognition and extraction ability of street trees by improving the RandLA-Net network. In the data preprocessing stage, the ground points are first removed by calculating the point cloud slope value, and the outliers are removed by statistical filtering algorithm. Subsequently, two key improvements were made to the RandLA-Net network: first, the cross-entropy loss function in the original network was replaced by the focus loss function to reduce the impact of the imbalance in sample size on network segmentation, and to enhance the sensitivity of the model to samples that are difficult to classify; Second, the Efficient Channel Attention (ECA) module is introduced to strengthen the model's ability to learn and express street tree features. In order to verify the effectiveness of the proposed improved method, this study uses the road dataset of Yongchang Road in Beijing to carry out comprehensive training and testing. Experimental results show that the improved model has significant improvement in segmentation accuracy and Intersection over Union (IoU), which fully proves the effectiveness of the proposed improved strategy.