Machine Learning Approaches to Multi-Sensor Fusion: Overcoming the Limitations of Extended Kalman Filters
Jimeng Yang · 2025
In recent years, multi-sensor fusion for autonomous driving has become a hot topic in the automotive industry. Past researches have mainly focused on using Extended Kalman Filters to process data. However, Extended Kalman Filters may produce large errors in specific environments with strong nonlinearity. In order to improve the accuracy of data simulation, this paper proposes to use machine learning to process multi-sensor data. This model uses IMU, GNSS, and LiDAR multi-sensor data as data sets. This paper compares the simulation effects of multiple machine learning models, and finally selects the optimal model LightGBM based on simulation accuracy and data processing time as evaluation criteria. Compared with the Extended Kalman Filter, the accuracy of trajectory estimation of this model has been significantly improved. Experimental results show that machine learning models have higher simulation accuracy than Extended Kalman Filters and have extremely high data processing efficiency.