SLOT-MPC: A Hierarchical Whole-Body Model Predictive Controller to Enhance Simultaneous Localization and Object Tracking for UAVs
Senior
AI summary
Problem
Current UAV tracking methods typically decouple active SLAM and active object tracking, causing performance degradation in GPS-denied, feature-scarce environments. There is no unified approach to simultaneously optimize both self-localization and dynamic object estimation.
Approach
SLOT-MPC employs a two-layer model predictive control architecture: an upper-layer controller plans optimal chasing paths to minimize object estimation uncertainty, while a lower-layer controller optimizes whole-body motion and camera viewing angles to maximize visual localization quality under object-visibility constraints.
Key results
- Integrates a time-continuous information filter into MPC for real-time object tracking
- Boosts visual localization accuracy by over 50% in featureless environments via sampling-optimization view planning
- Unifies trajectory and gimbal control in a hierarchical framework to simultaneously solve active SLAM and object tracking
- Validated through experiments with open-source implementation released
Why it matters
Enables robust, perception-aware autonomous UAV operations in challenging GPS-denied environments for surveillance and dynamic monitoring.
Abstract
This paper proposes SLOT-MPC, a hierarchical model predictive control framework for a system of multirotor Unm- maned Aerial Vehicle (UAV), which aims to minimize uncertainty in estimating both ego-motion and a moving object, thus enhanc- ing the performance of Simultaneous Localization and Object Tracking (SLOT). The framework consists of two layers: OT- MPC (Object Tracking-Model Predictive Controller) for high- level path planning, and a full model whole-body SL-MPC (Self Localization-Model Predictive Controller) for path tracking and view control. The OT-MPC uses a point-mass model with a pro- posed time-continuous information filter to minimize object esti- mation uncertainty and computes optimal chasing paths online in a receding horizon manner. Subsequently, to improve visual-based self-localization, the SL-MPC is developed to track the path gener- ated by the OT-MPC, while simultaneously optimizing perception objectives considering observed features to improve visual-based localization. Thus, optimal control sequences for the aerial vehicle are obtained in real time. Experiments are performed to validate the practicability of our approach. We will release our implemen- tation as an open source package for the community.