CoD-Pred

Currently, the limitation of publicly available multi-drone collaborative perception datasets lies in their focus on detection and segmentation tasks, with minimal support for trajectory prediction. To address this gap, we introduce the CoD-Pred dataset, a simulated dataset specifically designed for multi-drone collaborative trajectory prediction using the CARLA simulator. CoD-Pred employs four drones with identical equipment, including the same communication devices, sensors, and computing units. The drone swarm is positioned at intersections or along roadsides, enabling observation of the same traffic flow from different angles. The drones maintain a constant flight altitude of 75m, capable of covering a monitoring area of approximately 200 m × 200 m. The optical sensors carried by the drones have a sampling frequency of 2 Hz and a resolution of 1600 × 900. Each drone records sequences lasting 20 seconds, comprising 64,000 synchronized images.CoD-Pred contains 200 scenarios, divided into 170 training scenarios and 30 testing scenarios, including annotations for 24,768 vehicle instances and 4,851 pedestrian instances. Following the NuScenes data annotation approach, the dataset records both 2D and 3D labels for each instance, supporting not only basic perception tasks but also instance segmentation and trajectory prediction tasks. The CoD-Pred dataset embodies the inherent multi-faceted challenges of remote sensing scenarios, such as long-range perception, small object detection, and target occlusion.