A Multi-sensor Data Fusion Method for Vehicles Based on Improved GNN Association

Intelligent driving technology has become a hot direction in the automotive industry. Accurate and stable tracking and fusion of target information from multiple heterogeneous onboard sensors are prerequisites for intelligent vehicles to make correct decisions. To reduce target measurement errors and improve the stability of perception results under the constraints of low computing power of vehicle microcontrollers, this paper proposes a framework for multi-sensor data fusion in vehicles. Addressing the problem of temporary target disappearance due to sensor occlusion and the presence of irrelevant signals and noise, a single-sensor target tracking method based on Kalman filter and the lifespan approach is designed. To overcome the issue of erroneous associations commonly encountered in traditional global nearest neighbor (GNN) methods, an improved GNN data association method is proposed. It involves designing a target correction method for single-sensor tracking, followed by the design of a tracking gate setting method. The association cost is used for secondary target selection within the gate, and a weighted similarity function is defined to establish target associations. Finally, a convex combination fusion algorithm is used to fuse data from multiple sensors.