Point Cloud Multi-view 3D Target Detection Method Based on Complementarity
Zixuan Huang,Changjiang Jiang,Xiaoming Luo,Tong Lin
DOI: https://doi.org/10.1109/CAC57257.2022.10055885
2022-11-25
Abstract:The 3D target detection work based on point cloud view only considers the conventional view projection method, which loses a lot of point cloud information, resulting in poor target feature extraction ability. To solve this problem, this paper proposes a multi-view 3D detection algorithm for point clouds guided by complementary characteristics, which makes full use of the complementary characteristics between perspective and bird's-eye views. This algorithm designs an efficient cross-fusion feature block, which can effectively solve the problem of difficult feature extraction in the 3D object detection method based on point cloud view. The algorithm framework consists of four parts: point cloud two-dimensional processing, multi-view feature map fusion module, RoI pooling, and detection head. Using the pre-region proposal and cross-fusion features of RPN, the key point features, convolution features, and fusion features are aggregated to generate better target feature regions. Finally, the prediction of object category, 3D bounding box and object orientation is realized in the detection head. Experiments show that the proposed method has good performance in terms of accuracy and real-time performance of multi-target detection such as pedestrians and vehicles.
Computer Science,Engineering