Track segment association via deep spatiotemporal motion feature contrast
Zhonghe Hu, Sihao Gao, Wenwu Chen, Dongsheng Yang · Journal of Physics Conference Series · 2025
Abstract Radar systems often encounter interrupted tracking of flight targets during detection missions. Traditional methods, such as filtering, prediction, and similarity measurement, require laborious parameter tuning, limiting their practicality. Existing deep learning approaches overlook the inherent similarity between old and new track segments and the dominance of historical track features, leaving room for improved performance in track segment association. This paper proposes a track segment association algorithm based on deep spatiotemporal feature contrast, consisting of four modules: preprocessing, feature extraction, assisted attention, and classifier. The preprocessing module maps track vectors into a high-dimensional space using a multilayer perceptron and positional encoding. The feature extraction module employs a self-attention-based encoder to capture spatiotemporal motion features. The assisted attention module leverages BiLSTM and self-attention to extract temporal dependencies from old tracks and refine new track features via cross-attention. The classifier module computes association probabilities between track segments. Simulations demonstrate that our method significantly outperforms state-of-the-art techniques in track segment association accuracy.