Intent Recognition for Multi-agent Systems with Incomplete Information via Spatiotemporal Feature Fusion

This paper constructs six mathematical intention models for adversarial scenarios, such as encirclement and interception. A unified target point generation transforms intent recognition into time-series classification using a three-layer architecture for environment cognition, feature extraction, and inference. Kalman filtering with a nonlinear fade-in/out mechanism improves trajectory prediction and adaptability. Situation maps based on artificial potential fields integrate obstacles and agent dynamics. LSTM and Transformer networks extract multi-scale temporal and global features. Experiments on a multi-intent simulation dataset demonstrate the method’s high accuracy, robustness, and fast response, supporting intelligent reasoning and tactical decision-making in multi-agent systems.