Terrain-Aware Outdoor Navigation Using Gated Decision-Level Fusion and a Low-Cost Global Navigation Satellite System Based Extended Kalman Filter

초록

Outdoor ground-vehicle autonomy often degrades on rough terrain, where unreliable wheel odometry and degraded visual cues induce localization drift. To enhance robustness using low-cost sensors, this study presents a curvature-adaptive Extended Kalman Filter (EKF) combined with decision-level LiDAR-camera fusion for terrain-aware planning. The EKF prediction model switches online between constant-velocity (CV) and constant-turn-rate-and-velocity (CTRV) dynamics based on a curvature indicator, while smoothly blended process noise prevents covariance discontinuities. Localization updates rely exclusively on physically measurable signals, including low-cost GNSS, wheel odometry, and IMU data, with Mahalanobis gating and one-to-one association ensuring reliable observation selection. Separately, LiDAR-derived elevation and variance features and camera detections are fused via weighted least squares to produce a traversability-aware planning signal. This decision-level fusion output is not used to directly update the EKF state; instead, it guides path planning toward smoother trajectories that indirectly improve localization stability by reducing vibration-induced disturbances. IMU-based vertical-acceleration statistics drive adaptive odometry covariance inflation, mitigating overconfidence on uneven ground. All sensor streams are time-aligned to a unified update instant with latency-aware covariance inflation to preserve real-time performance. Experiments conducted on a Jetson-based platform equipped with an OS0-128 LiDAR, a ZED2i stereo camera, an IMU, and a low-cost GNSS system demonstrate sub-meter tracking accuracy on a flat S-shaped course (mean error 0.43 m, maximum 0.56 m against RTK ground truth). Vibration analysis further confirms over 50% reduction in high-magnitude acceleration peaks when terrain-aware planning is enabled.

키워드