Multi-convolution Module Fusion Network for Object Detection Based on 3D Point Cloud

The existing lidar data has the characteristics of sparsity and rotation invariance, which affects the effect of the algorithm in downstream tasks.Aiming at the problem of sparse point cloud data target detection, this study improves the PartA2 model and proposes a Multi-convolution module fusion network. By improving the Refconv module and combining the EMAC module, a serial structure backbone network is proposed to optimize the feature extraction effect of the original backbone network. Through experiments, the private data set in the garage environment has good visual performance. For vehicles, pedestrian 3D indicators are increased by 1.2 % and 1.5 % respectively compared with the original model. The BEV index of bicycle objects with high detection difficulty is increased by 1.3 %, and the 3D index is increased by 1.6 %.