SRPT: State Refinement ProbSparse Transformer for Tactical Intention Recognition of Cluster Targets

Tactical intention recognition (TIR) is a crucial aspect of military situation awareness. Traditional methods for TIR, such as template matching and Bayesian networks, are challenging to accurately capture the spatiotemporal characteristics of cluster targets. In order to address this issue, we propose the state refinement ProbSparse transformer (SRPT) based on deep-learning method. SRPT network mainly comprises three distinct components, namely cluster-shared temporal embedding, ProbSparse transformer encoder, and state refinement (SR) layer. The first part enhances the input with temporal positional encoding, and adds a learnable token to each sequence data to represent global information. Then, we use ProbSparse self-attention to replace vanilla self-attention, which can precisely extract temporal features while reducing computational complexity. With the aim of obtaining spatial mutual information among cluster targets, SR layer relies on a message passing mechanism, which assigns higher weights to neighboring ones with greater similarity. So SRPT can refine current states of all targets within a cluster and capture spatiotemporal features for TIR. We build a tactical intention dataset and conduct experiments on it. The results show that our method SRPT accurately predicts the tactical intention of targets in land battlefield scenarios, surpassing all compared models.