Multisource Data Fusion and Algorithm Comparison in Understanding Driving Intent

This study proposes an integrated framework for detecting a vehicle's driving intention using multi-source data fusion, thereby addressing the limitations of relying on single data sources in complex traffic scenarios.Environmental perception data (other cars, pedestrians, traffic lights) are processed via YOLOv8 and BEV to extract fine-grained spatial features; vehicle motion data (speed, acceleration, steering angle) are denoised with Kalman filters to eliminate sensor drift and road unevenness interference; driver behavior insights from eye-tracking (gaze points, blink frequency) and HMI (button operations, voice commands) derive regularized features reflecting attention and intentions.The framework evaluates three algorithms: rule-driven methods offer interpretability but low accuracy, statistical learning boosts precision but depends on manual features, and deep learning achieves high accuracy yet faces latency and black box issues.Single-modality approaches lack adaptability, while multimodal fusion reaches high accuracy but raises costs.To balance trade-offs, it suggests cross-modal lightweight integration (reducing overhead) and federated learning (protecting privacy) for algorithm selection and autonomous driving deployment, guiding practical use.