A new dual-channel trajectory prediction model

In this paper, we propose an innovative dual-channel trajectory prediction model designed to enhance both the accuracy and robustness of trajectory predictions in complex dynamic environments. The model leverages advanced components including Convolutional Neural Networks (CNN), Efficient Channel Attention Networks (ECANet), and Bi-directional Gated Recurrent Units (BiGRU) to establish a highly efficient and resilient prediction framework. Specifically, the model features two parallel CNN channels, each independently extracting spatial and temporal features from the input trajectory data. This parallel structure not only strengthens the model’s feature learning capabilities but also captures the diversity and complementary information within the aerial trajectory data through distinct convolutional kernels and pooling operations. At the end of each CNN channel, we integrate ECANet, which enhances the model’s ability to focus on critical trajectory features while suppressing irrelevant information via its efficient channel attention mechanism. Following this, the two ECANet-optimized feature representations are combined and integrated into a more comprehensive feature vector using feature concatenation. This final feature vector is then passed into the BiGRU network for sequential trajectory prediction. The bi-directional nature of the BiGRU allows it to capture both forward and backward dependencies in the trajectory data, leading to more accurate predictions of the aircraft’s position, speed, and heading at future time steps. Experimental results demonstrate that the proposed dual-channel trajectory prediction model significantly outperforms existing methods in terms of both prediction accuracy and stability, as shown by evaluations on the ADS-B real dataset.