Bedside Depth Simulations: rendered bed exits, falls and look-alikes seen by a depth camera in front of a hospital bed

Bed exits, falls and look-alike movements, seen by a depth camera on the wall in front of a hospital bed. 62 rendered 3D clips with depth for every frame and exact ground truth, and 52 AI-generated hospital videos from the same camera position, with simple detection rules as a baseline. Free to use under CC BY 4.0; citation of this record is required. Why. A depth camera at the foot of a hospital bed sees the whole bed and the floor beside it. Of the 53 public datasets we checked, none shows real patients leaving hospital beds from there. From this position a lying patient looks upright in the picture, so testing a detector needs depth, exact event times and true body positions. 3D simulations (62 clips, R01 to R62). Rendered in Blender 4.5 with MakeHuman / MPFB bodies (CC0), seen by a virtual camera with the fields of view of a Luxonis OAK-D Pro W. Each clip gives: a colour video and a grey infrared-like video; a 16-bit depth image per frame (millimetres; 7,217 frames in all); exact event start and end times; per frame, 24 body points in 3D and in pixels, with visibility flags, hip height and trunk angle; both cameras' calibration. They show sitting up, leaving the bed, eight kinds of fall, and look-alikes: turning, lying back down, sitting on a chair, a blanket or bag on the floor, a nurse kneeling, putting on slippers and walking away. Two bodies (a man of about 74, a woman of about 79) appear, with seven camera positions in the first set. 22 clips also give the empty room's depth. AI-generated videos (52). 12 start pictures, drawn from text by Seedream 5.0 Pro, were animated by Wan 2.2 (32 videos), Veo 3.1 Fast (16) and Seedance 2.0 (4), through fal.ai. They are released exactly as made, with each video's prompt, settings, seed and a camera-stillness measure. They have no depth. Their labels (the scenario each prompt asked for) have not been checked against the videos, and in 17 of them the camera zooms or shifts. Baseline. Detection rules, each locked before the clips that judge it were made, with the body points they read (RTMPose-m with YOLOX-m) and their results, which a script reproduces exactly. Read before use (docs/KNOWN_ISSUES.md): the depth is ideal; a pinhole lens shows people about 45% smaller than the real lens would; fallen bodies are propped on one hand in 17 clips; the camera pose is known exactly, where a real camera must estimate it; the motion is hand-keyed, in one room, with two bodies. No real people and no patient data. Files. bedside-depth-sim_v1.0_data.zip: the 3D simulations. bedside-depth-sim_v1.0_generated.zip: the AI-generated part. The same AI-generated videos (b1_*.mp4, b2_*.mp4) and start pictures (SF*.png), also as single files, to watch or download one at a time. bedside-depth-sim_v1.0_repository.zip: documentation, metadata, baseline and code. Format guide and code: github.com/shahabafshar/bedside-depth-sim. Licence. Data: CC BY 4.0. Citation required: the credit this licence requires is a citation of this record, in any publication, presentation, product or shared copy that uses it. For the AI-generated files, also say they are AI-generated and name the model. Code: Apache-2.0, except the Blender scripts, which are GPL-3.0-or-later.