Synthetic LiDAR Dataset for Human Fall Detection

This dataset contains synthetic LiDAR point-cloud sequences representing human fall and non-fall activities. The dataset was created to support research on privacy-preserving fall detection systems for Ambient Assisted Living (AAL) environments. Because collecting real fall events with human participants is difficult and ethically constrained, the dataset was generated using a simulation-based approach combining Blender 3D and Mixamo human motion animations. A virtual LiDAR sensor was simulated using the Range Scanner add-on to capture point-cloud data for each animation frame. The dataset contains 1000 activity sequences, each consisting of 60 LiDAR frames, with approximately 128 LiDAR points per frame. Dataset Statistics Property Value Total sequences 1000 Fall sequences 500 Non-fall sequences 500 Frames per sequence 60 LiDAR points per frame ~128 Total frames 60,000 Total LiDAR points ~7.68 million Dataset Directory StructureDataset (Blender+LiDAR) 1000poses│├── Fall Data 500 poses│ ├── 50FallData_PoseSet part 1│ ├── 50FallData_PoseSet part 2│ ├── ...│ └── 50FallData_PoseSet part 10│├── No Fall Data 500 poses│ ├── 50 Drinking 1│ ├── 50 Running 2│ ├── 50 Sitting 3│ ├── 50 Squat 4│ ├── 50 Standing 5│ ├── 50 JumpDown 6│ ├── 50 Walking 7│ ├── 50 PosesLeftTurn 8│ ├── 50 Pushing 9│ ├── Stretching 10│ ├── running.blend│ └── stretching_code.blend Fall Dataset The "Fall Data 500 poses" directory contains simulated fall activities. Fall Data 500 poses│├── 50FallData_PoseSet part 1├── 50FallData_PoseSet part 2├── ...└── 50FallData_PoseSet part 10 Each subset contains 50 pose sequences: 50FallData_PoseSet part 1│├── Pose_000_OriginalPose├── Pose_001├── Pose_002├── ...└── Pose_049 Thus: 10 subsets × 50 poses = 500 fall sequences Non-Fall Dataset The "No Fall Data 500 poses" directory contains daily human activities. No Fall Data 500 poses│├── 50 Drinking 1├── 50 Running 2├── 50 Sitting 3├── 50 Squat 4├── 50 Standing 5├── 50 JumpDown 6├── 50 Walking 7├── 50 PosesLeftTurn 8├── 50 Pushing 9└── Stretching 10 Each activity folder contains 50 pose sequences. 10 activities × 50 poses = 500 non-fall sequences Pose Folder Each pose folder represents one activity sequence. Example: Pose_000_OriginalPose Each pose contains 60 LiDAR frames: Pose_000_OriginalPose│├── frame_01_frame_1.csv├── frame_02_frame_2.csv├── frame_03_frame_3.csv├── ...└── frame_60_frame_60.csv Thus: 1 pose = 1 activity sequence1 sequence = 60 LiDAR frames CSV File Format Each CSV file represents one LiDAR scan frame. Example file: frame_44_frame_44.csv Each row corresponds to one LiDAR point. Columns included in each CSV file: Column Description categoryID Object category identifier in simulation partID Human body part identifier X X coordinate of the LiDAR point Y Y coordinate of the LiDAR point Z Z coordinate of the LiDAR point distance Distance from LiDAR sensor X_noise X coordinate with simulated noise Y_noise Y coordinate with simulated noise Z_noise Z coordinate with simulated noise distance_noise Distance with simulated noise intensity LiDAR return intensity red Red color channel green Green color channel blue Blue color channel Features Used for Machine Learning In the associated research experiments, only the spatial coordinates were used: X, Y, Z Each activity sequence is represented as: 60 frames × 128 points × 3 coordinates Activity Types Fall Activities Forward fall Backward fall Side fall Non-Fall Activities Drinking Running Sitting Squat Standing JumpDown Walking LeftTurn Pushing Stretching Data Generation Methodology The dataset was generated using the following pipeline: Human character imported into Blender Motion animations applied from Mixamo Pose variations generated by rotating body joints. A virtual LiDAR sensor simulated using the Range Scanner add-on Each frame exported as a CSV point cloud Pose variations were created by applying controlled rotations to: arms legs torso head This produced diverse body configurations for each activity. Blender Files The dataset also includes some Blender project files (in the Non-Fall Dataset): running.blendstretching_code.blend These allow researchers to reproduce the LiDAR simulation, modify animations, generate additional synthetic data. Limitations The dataset is synthetic, generated in simulation. Real-world LiDAR sensors may introduce noise, occlusions, and environmental variability. Future work should validate models using real LiDAR hardware. Intended Use This dataset can be used for research in: Fall detection Human activity recognition LiDAR sensing Privacy-preserving monitoring Deep learning for healthcare Ambient Assisted Living (AAL) Contributors Amir Ali — Dataset design, LiDAR simulation, data generationAhmad Alsharoa — Research collaboration, Research supervisionTeodoro Montanaro — Research supervisionLuigi Patrono — Research supervisionIlaria Sergi — Research collaboration