UB-SuitFall: A two-view CCTV fall detection dataset recorded in an age-simulation suit

UB-SuitFall is a two-view CCTV fall detection dataset in which two young adults wearing an age-simulation suit performed five types of simulated fall and five activities of daily living (ADLs), three trials each. Every trial was recorded simultaneously by a frontal and a lateral CCTV camera, and every fall is annotated separately for each camera with its start, impact and end frame. Summary- Participants: 2 (S01, S02), male, aged 20 and 23, wearing the Sakamoto Aged Simulation Suit II (model M176, size L) with a 0.5 kg weight in every weight pocket; goggles, ear plugs, gloves and belt covers were not worn- Activities: 5 falls (forward fall using hands, forward fall using knees, falling backwards, falling sideward, falling while sitting in an empty chair) and 5 ADLs (walking, standing, sitting, picking up an object, lying down), 3 trials each- Cameras: 2 fixed EyeSec CCTV cameras, CCTV-A (frontal, about 1.90 m from the mat) and CCTV-B (lateral, about 2.10 m), both about 1.65 m high; 1920x1080 at 30 fps, not frame-synchronised- Recordings: 108 camera-specific recordings from 54 trials (30 falls, 24 ADLs)- Frames: 31,239 frames at 10 fps (52.1 min), 5.0% labelled FALL- Recorded on 15 July 2026 at the Faculty of Health Sciences (FIKES), Universitas Brawijaya, Malang, Indonesia Files- UB-SuitFall_S01.zip and UB-SuitFall_S02.zip: one folder per participant containing CCTV-A/ and CCTV-B/: videos, named ActivityXTrialY.mp4 Frames/CCTV-A/ and Frames/CCTV-B/: frames at 10 fps, ActivityXTrialY/00001.jpg, 00002.jpg, ... fall_labels.csv: one row per recording, with activity, category (FALL or NON_FALL), number of frames, and the fall start, impact and end frames for that camera frame_labels/: one CSV per recording (e.g. S01_A01_T1_CCTV-A.csv) with columns frame, filename, state (PRE_FALL, FALL, POST_FALL or NON_FALL), label (1 for FALL, 0 otherwise) and label_3class (0 normal, 1 falling, 2 after the fall)- MANIFEST.csv: one row per released recording, with the paths of its video, frames and label file Frame numbers are 1-based and match the frame file names. Unzip both participant archives into one folder next to MANIFEST.csv to obtain the layout the manifest refers to. AnnotationFall start is the first frame of the loss of balance, impact is the first frame at which the pelvis or trunk is stopped by the ground, and fall end is the first frame at which the body comes to rest. Candidate frames were proposed automatically from the hip trajectory of an off-the-shelf pose model, and one annotator checked and corrected every candidate. Frames from fall start to fall end are labelled FALL; all frames of ADL recordings are NON_FALL. Missing recordingsOf 120 planned recordings, 108 are available. The 12 unavailable recordings are all from S01 (standing trial 2, picking up an object trials 2 and 3, lying down trials 1 to 3, each from both cameras), because the videos were missing or corrupted. S01 therefore has no lying-down recordings and only one picking-up trial. These recordings are listed in fall_labels.csv with n_frames = 0. Code and baselineThe code used to extract frames, annotate falls and assemble the release, and a leave-one-subject-out baseline benchmark (Random Forest, SVM and LSTM on suit-silhouette features), is available at https://github.com/BaboonBongo/UB-SuitFall. In the baseline, Random Forest reached a window-level F1 of 0.785 and a PR-AUC of 0.935, and detected all 60 camera-specific fall instances with 14.3 false alarms per hour of ADL recording. With two participants, these results are an initial reference rather than a definitive comparison. PrivacyBoth participants consented to the publication of their recordings, including identifiable images of their faces. Other people, not wearing the suit, are sometimes visible in the background. Please respect the privacy of everyone shown and do not attempt to identify them or link the recordings to other information about them. LicenceCreative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). ContactMuhammad Ali Fauzi ([e-mail]), Faculty of Computer Science, Universitas Brawijaya.