To meet the high-precision requirements of perception systems for autonomous driving and robotics

FusionDetect

a two-stage 3D object detection method that fuses LiDAR and camera sensors

is proposed to address the long-distance object recognition challenges of existing LiDAR detectors due to the sparsity of point clouds.This method innovatively fuses the geometric information of the LiDAR with the texture information of the camera

employs region of interest(RoI)pooling technology to uniformly process multimodal data

and designs intra-modal self-attention and cross-attention mechanisms to achieve feature enhancement and information fusion.Experiments show that on the KITTI and Waymo benchmarks

FusionDetect not only achieves a breakthrough improvement in the performance of mainstream detectors

but also surpasses the baseline model with a 6.45% mean average precision(mAP)improvement on the Waymo dataset

significantly outperforming existing two-stage methods.These results validate the key role of multimodal sensor fusion in improving the accuracy of 3D object detection.