MonoSPDC：A Novel Monocular 3D Object Detection Using Global-Channel Atten- tion Optimization

Abstract

With the development of autonomous driving and mobile robotics, monocular 3D object detection has attracted attention in the above fields due to its advantages of low cost, fast speed and low power consumption. However, both mainstream monocular 3D object detection network structures currently face difficulties in practical application in these fields. One issue is the low accuracy of convolutional network structures. Another issue is the complex structure, high computational complexity, large training dataset, and LiDAR assisted training problem of the Transformer network structure. To solve the above problems, a novel monocu-lar 3D object detection algorithm named MonoSPDC is proposed in this paper. In MonoSPDC, we propose the Feature Global-Channels Recombination Module (FGCR). FGCR uses the attention mechanism based on convolutional structure to perform attention operations on the global and channel dimensions of the feature map. This can make the feature map more distinct, while also making the generation of heatmap more reliable and accurate. we also propose the Multi-View Predictive Integration module (MVPI) in MonoSPDC. MVPI can realize the uncertain fusion of 3D corner information based on 2D feature regression and depth information based on depth feature map, so as to improve the accuracy of the final depth estimation. In summary, proposed Mono-SPDC does not require other additional auxiliary training data. It can retain the speed advantage of the convo-lutional structure while achieving the accuracy of the Transformer structure. In the test on the public dataset KITTI, our proposed MonoSPDC has achieved very excellent results.