Maneuvering Targets Fusion Filtering for Vehicle-Mounted Multisource Heterogeneous Sensor Using MHS-GRU-Fusion Net

Abstract

Autonomous driving requires environmental perception and target tracking with vehicle-mounted heterogeneous sensors. The data from those devices demand fusion and filtering techniques, operations which are challenged by a lack of proper solutions for model and time mismatches. Traditional target tracking methods based on Kalman filtering suffer from model mismatch under maneuvering targets, while time alignment of asynchronous multisource sensors via prediction inevitably introduces time registration errors, degrading tracking accuracy. To address these issues, this article proposes a multisource heterogeneous sensor (MHS) fusion network that combines the fusion filtering algorithm with gated recurrent units (GRUs). The MHS-GRU-Fusion Net implements maneuvering target tracking by integrating multiple GRUs and fully connected (FC) layers for state prediction, updating, and covariance propagation. By establishing an explicit correspondence between the MHS tracking formulation and the deep network architecture, the proposed network preserves the computational logic of traditional multisensor target tracking and effectively mitigates model mismatch and time registration errors.