Object perception based on bird's-eye view(BEV) is one of hot issues, but studies on multi-sensor fusion for BEV are still insufficient. Therefore, a multi-sensor fusion module based on spatial-channel attention is proposed. Spatial errors between multiple sensors can be effectively corrected by adding local attention mechanisms to features of different modalities. By using transpose attention operations, the image and point cloud data are fully integrated to resolve the heterogeneity between different modal semantics. Consequently, the fused BEV features achieves more comprehensive and accurate perception by effectively combining the unique information of each sensor without introducing spatial misalignment. Experiment on nuScenes dataset and extensive ablation experiments show that the proposed fusion module effectively improves the accuracy of object detection. Visualization results demonstrate that the fused features can capture more complete and accurate information, especially in distant objects detection.

The spatial consistency of multi-source sensing data is established as the foundation for the fusion of roadside multi-modal data, playing a crucial role in vehicle-to-infrastructure and roadside intelligence. However, existing roadside multi-modal datasets predominantly focus on recognition tasks such as object detection, lacking various spatial transformation information between multi-source sensors. This deficiency hinders the research into the spatial consistency problem of multi-source data at the roadside. Therefore, a dataset specifically designed for the study of the spatial consistency problem in roadside multi-source data is constructed in this paper-InfraCalib(https://github.com/chenzhiwei888/InfraCalib-Dataset). The dataset comprises over 230,000 frames of images and point cloud data, collected by two roadside smart mobile devices, covering diverse changes in scenes, modalities, lighting, device spatial positions and sensor postures. By matching feature key point pairs to correlate multi-modal data, a perspective-n-point(PnP) problem is constructed, and the extrinsic parameter matrix is solved using the minimum reprojection error method, serving as an approximate ground truth label. Finally, an experimental analysis of the classic feature matching algorithm is conducted on the InfraCalib dataset, and the discussion revolves around the quantitative evaluation indicators for the calibration of external parameters in multi-source sensors.