Adaptive Vision–ToF Sensor Fusion for Fault-Tolerant Spud Height Estimation in Autonomous Dredging Systems

Accurate and reliable spud height estimation is fundamental to autonomous cutter suction dredging, where conventional sensors may be affected by harsh marine conditions, degradation, and single-point failures. This paper presents a fault-tolerant vision–ToF sensing framework combining a stereo-camera subsystem, discrete Time-of-Flight (ToF) reference-event detection, linear Kalman filtering, adaptive measurement weighting, and online two-point calibration. A laboratory platform implemented on an NVIDIA Jetson Orin Nano (Santa Clara, CA, USA) was evaluated using eight programmed software-injected measurement faults, with repeat trials for camera_noise and camera_dropout. For the three reference-event-path fault scenarios, peak estimator-reported position variance remained below the baseline peak of 0.87 px2, whereas camera-side degradation caused substantially larger covariance increases, reaching 85.21 px2 in a camera_dropout repeat. Across four camera_noise repeats, adaptive weighting reduced the mean absolute frame-to-frame output change by 17.3% relative to fixed-covariance replay. The complete online pipeline achieved a mean frame-processing interval of 80.9 ms (approximately 12.4 frames/s) across six repeat experiments. Because no continuous independent dynamic ground truth was available, the results characterise estimator-reported uncertainty and tracking behaviour rather than absolute physical position accuracy. The findings indicate that vision is the primary continuous position channel, while ToF provides discrete TOP/MID physical reference events supporting calibration and event-triggered correction.