Depth-Augmented Pose Tracking for Autonomous Robots: A Filter-based Multi-Sensor Fusion Approach

Abstract

Robotic platforms operating in GPS-denied environments require robust ego-motion estimation systems that fuse complementary sensor modalities under onboard computational constraints. This paper proposes a navigation framework estimating six-degree-of-freedom (6 DoF) robot pose in unstructured scenes using a monocular camera stream, inertial measurement unit (IMU) data, and sparse depth cues within the multi-state constraint Kalman filter (MSCKF) architecture. The key innovation integrates 3D landmark measurements into visual feature tracks, reducing positional uncertainty and drift accumulation compared to vision-only approaches. The method is efficient enough for resource-constrained systems such as micro aerial vehicles and small ground robots. The measurement fusion strategy is analytically derived and evaluated on aerial robot trajectory datasets. Results show improved tracking accuracy and stability in challenging indoor and outdoor scenarios without GPS, enabling prolonged autonomous missions in complex 3D environments with real-time pose feedback and low computational burden.