A Multi-Scenario Dataset for Long-Term Indoor Localization and Pedestrian Behavior Analysis in Dynamic Environments

This dataset provides multi-scenario recordings for evaluating long-term indoor localization and pedestrian behavior analysis in dynamic environments. All data were collected in a 40×10 m entrance hall at Leibniz University Hannover, using a mobile multi-sensor platform and four synchronized overhead cameras.

The same space was recorded several times while the environment was intentionally changed: structural modifications, movable walls, furniture rearrangement, and natural pedestrian activity. These controlled variations allow direct comparison of algorithm performance under different levels of occlusion, scene change, and human motion.

The dataset supports research on LiDAR-based localization and mapping, sensor fusion, dynamic-scene handling, pedestrian tracking, behavior analysis, and human–robot interaction. Each sequence includes LiDAR data, multi-view camera recordings, and ground-truth trajectories.

Scenarios

The same physical space was recorded in three configurations:

1) Extreme occluded: Eight poster walls were arranged to create a narrow corridor. Most original structures are hidden from the LiDAR sensors, and pedestrians introduce additional occlusions.

2) Semi occluded: Several L-shaped poster walls, tables, and chairs partially block structural features but leave large parts of the room visible.

3) Free space: No artificial obstacles were added. Around 15 pedestrians moved naturally, causing only short-term occlusions.

Figure 1: Illustration of the three recording scenarios: extreme occluded, semi occluded, and free space. The left side shows the floor plan with sensor placement, pedestrian distribution, and ground-truth trajectory, while the right side provides sample views from the four fixed cameras.

Dataset Statistics

Eight sequences were recorded: 2 extreme occluded (S0–S1), 3 semi occluded (S2–S4), and 3 free space (S5–S7). Each sequence includes LiDAR data, synchronized overhead camera recordings, and ground-truth data.

Sensor Platform

The mobile platform is a push-cart with a rigid frame carrying the following sensors:

• Ouster OS1 LiDAR
• Hesai Pandar64 LiDAR
• Ricoh Theta X 360° camera (Not part of the Dataset)
• IMU (Not part of the Dataset)

The LiDAR data is recorded via ROS 1 as rosbag files on an onboard mini PC.

Data Structure

The LiDAR data is present in a single compressed rosbag file per sequence. The file contains 3 topics:

• /hesai/pandar
  • topic type: sensor_msgs/PointCloud2
• /ouster/imu
  • topic type: sensor_msgs/Imu
• /ouster/points
  • topic type: sensor_msgs/PointCloud2

Overhead Camera System

(Work in progress)

Four fixed cameras mounted in the corners of the hall record synchronized RGB video:

Resolution: 640×480 px

AI processing: Raspberry Pi 5 + Sony IMX500 edge device running YOLOv8 for pedestrian detection

Recording mode: Frames are saved only when pedestrians are detected to reduce storage load

The four cameras are timestamped with synchronized NTP clocks and GPS-based UTC time.

The images can be accessed via the following link: https://drive.google.com/drive/folders/1-gkseQ_2L3_ElLoS5NNMoO2YngLv-X0-?usp=sharing

Figure 2: Examples of pedestrian detection from the four fixed cameras under varying viewpoints and occlusion conditions.

Work in progress: data might be changed