F-ResPoint: a multi-modal fusion object detection algorithm based on frustum culling

Accurate and lightweight 3D object detection is crucial for the perception ability of mobile intelligent robots. Aiming at the problem that current 3D object detection is difficult to balance lightweight and accuracy, this paper presents a multimodal fusion object detection method, F-ResPoint, based on point cloud culling in vision frustum. F-ResPoint is a lightweight two-stage object detection method, in which the frustum point cloud generation algorithm based on the improved YOLOv5 is used first to obtain the point cloud containing the object information in the view frustum from the point cloud in the original field. Then, the point cloud detection network based on the Residual-SA module is utilized for detecting the three-dimensional (3D) object from point cloud in the view frustum and obtaining its position and actual size of the object in 3D space. Experimental results demonstrate that this method achieves high detection accuracy and maintains good real-time performance while remaining lightweight.