Roadside Lidar Object Detection Based on Improved Second

With the continuous progress of deep learning technology and the growing affordability of LiDAR, research on 3D object detection technology based on LiDAR has attracted much attention in the field of vehicle-infrastructure cooperation. Currently, deep learning algorithms for 3D object detection based on LiDAR are predominantly utilized in autonomous driving data collected by vehicle-mounted LiDAR. However, fewer studies were focused on deep learning-based object detection using roadside LiDAR data. This paper proposes a 3D object detection algorithm based on roadside LiDAR, named DE-SECOND, which is an adaptation of the well-established SECOND network. DE-SECOND employs enhanced dynamic voxel coding to overcome the limitations of conventional voxel coding, which often discards valuable information when the number of points and voxels exceeds the preset threshold. Additionally, in order to allow sparse convolution to more effectively ascertain the significance of various positions, focal sparse convolution layers are introduced into the DE-SECOND backbone network. Finally, to render the detection head more perceptive and attuned to the task, a dynamic detection head is proposed. DE-SECOND has undergone training and evaluation on the DAIR-V2X-I dataset. The empirical results indicate that DE-SECOND outperforms SECOND, boosting vehicle detection accuracy by an average of 2.77% and pedestrian detection accuracy by an average of 4.6%.