Deep Learning for Tree Segmentation in MLS Data

Street tree inventory is an important part of urban forest inventory. Mobile laser scanning (MLS) technology has a strong data acquisition ability inside the canopy and the trunk and is suitable for parameter estimation at the tree level. The first and key step is to segment individual trees from the street MLS data, which is a challenging problem due to the semantic gap. This paper proposes an efficient and effective method for street tree segmentation from MLS data using deep learning-based image instance segmentation.

By Q. Li, et al.

First, the three-dimensional (3D) street point cloud captured by the MLS system is mapped to a two-dimensional (2D) RGB image. Next, pixelwise tree proposals are segmented from the street image by a trained deep learning-based image instance segmentation model. Then, the 2D segmentation mask is mapped back to the 3D street point cloud to generate pointwise tree proposals. Finally, the proposals are optimized to obtain the final results.

To evaluate and verify algorithm performance, an MLS point cloud is collected from a 1481.8 m-long one-side urban street containing various objects. Three deep learning-based image instance segmentation algorithms, YOLACT, BlendMask, and YOLOv8, are carried out, and YOLOv8 achieves the best results in terms of both accuracy and speed. YOLOv8 has the highest segmentation accuracy, with IoU= 0.85:0.05:0.95, with an average segmentation time of 26 ms per image.

In the experiments comparing the algorithms to the existing hierarchical segmentation and classification segmentation methods, the proposed method outperforms the other two methods in accuracy and is faster. The precision is 0.9988, the recall is 0.9986, the score is 0.9987, and the time per scanline is 4.05 ms. Moreover, the proposed method can be applied to MLS data with a broad range of resolutions by introducing image resizing.

For the full paper CLICK HERE.