Multi-Level Rotational Equivariant Object Detection Network Based on BEV Fusion

With the development of autonomous driving systems, the detection of three-dimensional objects in road scenes has garnered widespread attention. However, most single-sensor or multi-sensor fusion-based object detection methods do not consider the synchronized rotation of the captured scene owing to vehicle movement in real road scenes, which impairs object detection performance. To address such problems, this study proposes a multi-level global rotational equivariant object detection network framework based on multi-sensor fusion to alleviate the difficulty of object detection caused by scene rotation and thereby improve object detection performance. First, the interior of the voxels is encoded by the distance between each point to enhance the local point cloud geometric information and extract the global rotational equivariant features of the voxels. Second, the semantic information of the image is introduced, and global rotational equivariant features are extracted to further improve the network performance. Finally, the point cloud and image information, all with rotational equivariants, are fused on a Bird's-Eye View (BEV) and embedded in a group equivariant network to extract the global rotational equivariant features on the fused BEV level. Experimental results on the nuScenes validation set show that the network architecture achieves a mean Average Precision (mAP) of 68.7% and a nuScenes Detection Score (NDS) of 71.7. Moreover, the mean Average Orientation Error (mAOE) decreases to 0.288. Compared with mainstream object detection methods, the proposed method realizes the rotational equivariance of the network architecture and improves performance. In addition, each component plays an important role in improving the object detection performance of the overall network architecture.