CoBEVFusion Cooperative Perception with LiDAR-Camera Bird's Eye View Fusion

Autonomous Vehicles (AVs) use multiple sensors to gather information about their surroundings. Connected Autonomous Vehicles (CAVs) share sensor data for increased safety and reliability through cooperative perception. However, most recent approaches in cooperative perception share unimodal information perceived using a single sensor such as only camera video data or only LiDAR point cloud data or perform multi-modal data fusion at the early or late stage. In this research, we explore vehicular perception utilizing intermediate fusion of mul-timodal camera video and LiDAR point cloud data. We propose the Dual Window-based Cross-Attention (DWCA) model which extracts and fuses selected camera and LiDAR data features and projects that onto a Bird's-Eye View (BEV) representation on a single ego vehicle. We demonstrate that using multimodal camera and LiDAR data, our DWCA model exceeds the performance of state-of-the-art (SOTA) models using unimodal data in vehicular object detection and segmentation tasks on a single vehicle. Next, we propose a model for Cooperative Perception, CoBEVFusion, which aggregates the fused BEV representations obtained from surrounding CAVs using a 3D Convolutional Neural Network. We validate our CoBEVFusion framework on the cooperative perception dataset, OPV2V, for two perception tasks: 3D object detection and BEV semantic segmentation. The CoBEVFusion model outperforms SOTA models in object detection tasks when using unimodal data or multimodal camera-LiDAR data.