METHODS OF ADAPTIVE FUSION AND COLLISION RESOLUTION OF BEHAVIORAL MODALITIES FOR HUMAN ACTION PREDICTION IN INTELLIGENT INFORMATION-CONTROL SYSTEMS

Abstract

This research presents an adaptive multimodal fusion framework for human intent recognition in intelligent information-control systems under asynchronous, noisy, and conflicting behavioral inputs. The proposed approach combines Dynamic Time Warping, Variational Autoencoders, Graph Attention Networks, Jensen-Shannon Divergence, Dempster-Shafer Evidence Theory, and an Extended Kalman Filter to align heterogeneous data streams and resolve cross-modal conflicts. Experimental evaluation was performed using a custom multimodal dataset containing 1500 control sessions and more than 450,000 temporal frames. The proposed framework achieved 96.4% accuracy under nominal conditions, 93.8% under noise-degraded conditions, and 91.2% under high-collision conditions, with an average inference latency of 23 ms and a false execution rate of 1.8%. Compared with baseline fusion architectures, the proposed approach demonstrated improved robustness and reliability under adverse operating conditions. The obtained results confirm the applicability of the developed framework to real-time multimodal human - computer interaction and intelligent control systems.