Pervasive Systems (PS)

Determining Sensor Priority for Environment Identification During Indoor-Outdoor Transitions

Determining Sensor Priority for Environment Identification During Indoor-Outdoor Transitions

Problem Statement

Reliable localization is essential for autonomous robots and vehicles operating across diverse environments. During transitions between open outdoor areas, urban streets, tunnels, underpasses, and covered passages, the usefulness and reliability of different sensors (camera, LiDAR, GNSS/INS, odometry) change significantly due to lighting variation, structural occlusions, limited sky visibility, and signal blockage.

Most localization and sensor fusion systems treat sensor contributions as static, without considering which sensors are more suitable for the current operational environment. However, for robust indoor–outdoor (IO) transitions, a robot must first recognize the type of environment it is in and then prioritize the most suitable sensors for localization accordingly.

Currently, there is limited research on which sensor modality is most informative for identifying the operational environment that directly influences localization performance. Understanding this relationship is an important step toward adaptive sensor fusion strategies for autonomous navigation.

This project investigates:

What is the priority and suitability of each sensor data type (Camera, LiDAR, GNSS/INS, Odometry) for identifying the operational environment: outdoor, indoor-like, or transitional?

The outcome will support the design of environment-aware sensor fusion strategies for autonomous systems.

Dataset

The project will use the Oxford RobotCar Dataset (Documentation), which provides synchronized:

  • Stereo camera images; LiDAR scans; GNSS/INS data and ground truth poses; Odometry data; Urban driving sequences containing open-sky roads, narrow streets, tunnels, and covered passages

Tasks

1.         Literature study

  • Indoor-outdoor transitions in robotics and autonomous driving.
  • Environment recognition using onboard sensors.
  • Sensor reliability in constrained and covered environments

2.         Dataset exploration and preparation

  • Explore selected sequences from the Oxford RobotCar dataset.
  • Define three environment classes:

◦             Outdoor (open sky, wide field of view).

◦             Indoor-like (tunnels, underpasses, covered or narrow structures)

◦             Transitional (entering or exiting such areas)

  • Design and implement a frame-level labeling scheme using:

◦             GNSS quality indicators

◦             Inspection of trajectories and surroundings

◦             Visual inspection of camera images and LiDAR

3.         Feature engineering and sensor processing through extracting environment-indicative features for each sensor modality:

• Camera

◦             Sky visibility

◦             Illumination/brightness characteristics

◦             Texture and structural cues

◦             Presence of vertical structures

• LiDAR

◦             Point cloud density and distribution

◦             Detection of vertical and horizontal structures

◦             Corridor or narrow-space characteristics

• GNSS/INS

◦             Signal quality variations

◦             Position noise or drift behavior

• Odometry

◦             Motion smoothness

◦             Changes in motion characteristics

4.         Model development

• Implement separate environment classifiers using features from:

◦             Camera-only data

◦             LiDAR-only data

◦             GNSS/INS-only data

◦             Odometry-only data

• Each model should predict whether the robot is in an outdoor, indoor-like, or transitional environment.

5.         Evaluation and analysis

  • Evaluate all models using standard metrics (accuracy, confusion matrix, F1-score).
  • Compare the performance of each sensor modality.
  • Analyze which features are most informative for environment identification.
  • Perform ablation comparison between sensors.
  •  Based on the results, rank the sensors according to their suitability for environment identification.

6.         Visualization

  • Visualize predictions along full driving trajectories.
  • Highlight regions where each sensor performs well or fails.
  • Provide clear graphical illustrations of sensor suitability across environments.

Reporting

• Document methodology, labeling strategy, feature extraction, models, and evaluation.

• Present comparative analysis and sensor ranking.

• Discuss limitations, potential improvements, and implications for adaptive sensor fusion in autonomous navigation.

Work Distribution

• 40% Theory: Literature review, understanding sensor behavior and environment characteristics.

• 45% Experiments & Modeling: Dataset preparation, labeling, feature extraction, model development, evaluation, and visualization.

• 15% Writing: Thesis report, figures, analysis, and discussion.

Contact

Alessandro Chiumento – Pervasive Systems, EEMCS, University of Twente (a.chiumento@utwente.nl)

Maya Aghaei – SMART Mechatronics and Robotics Research Group, Saxion University of Applied Sciences (m.aghaeigavari@saxion.nl)