Abstract
This paper presents a method for location estimation and obstacle tracking for indoor mobile robots using a laser scanner. Assuming a known map, the system matches laser measurements to estimate the robot's location via the Sequential Monte Carlo (SMC) method. Once localized, the system detects and tracks obstacles using a Joint Probabilistic Data Association (JPDA) filter, enabling the prediction of collision risks. Experimental results verify the effectiveness of the proposed approach.
Introduction
Autonomous indoor navigation requires solving two fundamental problems simultaneously: precise self-localization and reliable tracking of dynamic obstacles. While LIDAR sensors offer high precision, the challenge lies in efficiently processing large data volumes to distinguish between static environmental features (like walls) and moving entities (like humans). Traditional RF or vision-based methods often fall short due to indoor signal interference or lighting sensitivity. This research proposes a unified framework that leverages a known map to filter out static data, focusing computational resources on tracking dynamic obstacles.
Methodology
Localization via Sequential Monte Carlo (SMC)
The robot's position is estimated using a particle filter, or SMC method. The robot's state is represented by a set of particles, each representing a possible pose. As the robot moves, these particles are updated based on odometry and then weighted according to how well their predicted laser scans match the pre-existing grid map.
Particles with higher weights (better matches) are more likely to be selected during the resampling phase, causing the particle cloud to converge around the robot's actual location. This approach effectively handles sensor noise and provides a robust estimate compared to simple dead reckoning.
Obstacle Tracking via Joint Probabilistic Data Association (JPDA)
After localization, the system identifies laser returns that do not match the static map. These points are clustered using K-Nearest Neighbor (KNN) to represent potential obstacles. To track these clusters over time, the authors employ a JPDA filter.
The JPDA filter is particularly effective for multi-target tracking in cluttered environments. It uses a validation matrix ($\Omega$) to calculate the joint probability of various measurement-to-target associations. This allows the robot to maintain stable tracks even when obstacles move close to each other or when measurements are ambiguous. By estimating the velocity of these obstacles, the robot can predict future trajectories and proactively avoid collisions.
Experimental Results
The system was tested using a mobile robot equipped with a HOKUYO URG-04LX Laser Scanner in a laboratory setting.
- Localization Performance: The SMC method significantly improved accuracy. The average localization error was reduced from 49.77 cm using only wheel encoders to 18.79 cm with the SMC filter.
- Tracking and Safety: The JPDA filter successfully tracked multiple individuals moving at different speeds and directions. The predictive capabilities allowed the robot to differentiate between a person who would clear its path and one posing an immediate collision risk, leading to smoother navigation.
Conclusion
The integration of SMC-based localization and JPDA-based tracking provides a robust solution for indoor mobile robotics. By using a static map as a filter, the system achieves high efficiency in identifying dynamic obstacles. Future work will address limitations such as the 5-meter sensor range and the dependency on pre-mapped environments, potentially incorporating SLAM techniques and more advanced path planning that fully utilizes predicted obstacle trajectories.
Summary updated