3D Small Object Detection from Cameras and Point Clouds Using Five-Head Attention in a Fusion Method

In this paper, we focus on 3D point clouds, multimodal data fusion, and attention mechanisms. Through a survey of related research on 3D object detection based on multimodal data fusion, we identified three problems: (1) the detection accuracy of small objects, such as pedestrians and bicycles, is unsatisfactory; (2) the fusion training of two different models cannot match the efficiency of single model training; and (3) when there are long-range objects in the pseudo-image generated by the features, existing methods cannot maintain the original high accuracy, and the generalization ability of the model is weak. To solve these problems and improve the detection performance of single-modal based detectors, this paper introduces a new fusion network that mainly consists of a five-head attention module and a posterior decision fusion (CPFN) module. The five-head module suppresses noise interference by jointly considering channel, spatial, point, and voxel attention, while enhancing the understanding of key information about the object. Additionally, CPFNet uses a PointPillars network with an attention mechanism for decision fusion with a CascadeR-CNN network. Experimental results on the validation set of the KITTI dataset show that our proposed method far outperforms existing methods in the small sample category, whether compared to state-of-the-art fusion-based methods or point cloud neural networks.hods or point cloud neural networks.