Abstract
In this study, we propose an adaptive sensor fusion framework for real-time object detection in mobile robotic systems, designed to overcome challenges such as computational overload, inconsistent detection under dynamic motion, and redundant sensor processing. The framework dynamically integrates multi-sensor data (e.g., RGB cameras and LiDAR) using three coordinated mechanisms: (i) confidence-aware fusion, which selectively incorporates sensor measurements exceeding predefined detection-confidence thresholds; (ii) frequency-aware scheduling, which adapts sensor processing rates based on observed object motion to reduce unnecessary computation; and (iii) velocity-aware adjustment, which modulates fusion strategies according to the relative speed of moving objects. This approach ensures robust perception while optimizing resource usage. We validate the method through simulations comparing the adaptive fusion against a baseline fusion system under identical operating conditions. Performance metrics, including Real-Time Factor (RTF), Central Processing Unit (CPU) utilization, and memory usage, demonstrate that our method maintains RTF between 0.8 and 1.1, keeps CPU usage mostly below 9%, and stabilizes memory consumption around 15%, outperforming the baseline, which suffers from RTF drops and higher CPU loads due to indiscriminate processing. Detection accuracy is also benchmarked against the baseline, confirming reliable object recognition. These results indicate that the proposed method effectively balances computational efficiency and detection performance in environments with static obstacles, while its design incorporating confidence, frequency, and velocity-aware mechanisms lays the foundation for handling dynamic and visually degraded conditions, which are the focus of ongoing experiments.