Deep-Learning Model for Multi-Scale Construction Object Detection from 3D Sparse LiDAR

The construction industry requires an automated safety monitoring system to detect various objects during operations, thus preventing accidents caused by workers and large equipment. Utilizing LiDAR for monitoring, which can provide accurate 3D point cloud data, offers a promising solution. However, the objects involved in the safety monitoring have extremely different scales, such as large-scale excavators and small-scale workers, which pose challenges for detection with LiDAR. This study proposes a deep learning model that simultaneously detects complex multi-scale objects from LiDAR point clouds to address this issue. The proposed model generates object detection boxes using a voxel-based 3D sparse convolutional neural network and an anchor-free detection head based on candidate voxels. To improve the limited receptive field for detecting large-scale objects, additional feature extraction layers and improved large kernel convolution operations are incorporated into the network. A network with multi-scale 3D features and feature fusion from different scales is employed, enhancing the capability to detect objects of various sizes. The proposed method was trained and evaluated on the LiDAR data set we created. It achieves detection mAP exceeding 90% for both small-scale workers and large-scale excavators, indicating good multi-scale detection capability.