关键词: 三维目标检测, 自注意力机制, 特征融合, 动态图卷积, 小目标检测

Abstract: Aiming at the problem that the accuracy of detection of small objects such as cyclists and pedestrians in Three-Dimensional (3D) object detection is low, and it is difficult to adapt to complex urban road conditions, a 3D object detection network based on self-attention mechanism and graph convolution was proposed. Firstly, in order to obtain more discriminative small object features, self-attention mechanism was introduced into the backbone network to make the network more sensitive to small object features. Improve the ability to extract network features; Secondly, a feature fusion module was constructed based on the self-attention mechanism to further enrich the information of shallow network and enhance the feature expression ability of deep network. Finally, the dynamic graph convolution was used to convolve the boundary box of the prediction object to improve the accuracy of the prediction. This Network is tested on KITTI dataset, Compared to eight major networks such as TANet (Triple Attention Network) and IA-SSD (Instance-Aware Single-Stage Detector). The experimental comparison shows that among the 8 networks compared, the pedestrian detection accuracy of TANet and the cyclist detection accuracy of IA-SSD are the best. The pedestrian detection accuracy of the network in this paper is increased by 12.12, 13.82 and 11.03 percentage points compared with TANet under the three difficulties. The detection accuracy of riders is 3.06 and 5.34 percentage points higher than that of IA-SSD on medium and difficult points. In summary, the network proposed in this paper can be better applied to small object detection tasks.

Key words: Three-Dimensional (3D) object detection, self-attention mechanism, feature fusion, dynamic graph convolution, small object detection