High-granularity trajectory data of vehicles and pedestrians is essential for traffic simulation, behavior modeling, and intelligent transportation research. CrossTraj is a bi-season trajectory dataset captured at a midblock crosswalk in Milwaukee, WI, using synchronized panoramic roadside cameras covering 200 meters of roadway. Two one-hour recordings from winter and summer capture seasonal and traffic dynamics. The processing pipeline applies YOLOv9e for detection, DeepSORT for stable multi-object tracking, and monocular 3D estimation for depth, orientation, and dimensions. Detected objects are projected to a real-world ground plane via homography and unified across camera views through a stitching module. Kalman filtering and occlusion-aware smoothing reduce noise and fragmentation. The resulting dataset includes hundreds of continuous trajectories that reveal behaviors such as yielding, hesitation, and near-conflict interactions. Evaluations on identity continuity, alignment, and velocity confirm reliability. CrossTraj provides a benchmark for vehicle-pedestrian interaction studies, urban traffic modeling, and safety analysis, and is publicly available.
