MCK-Net: A Pedestrian Trajectory Prediction Network With MRF Clique Model and KAN

Pedestrian trajectory prediction within open social environments is a vital component of mobile robotics and autonomous driving. The inherent uncertainty and multimodality introduce considerable challenges for this task. Currently, an effective approach is to first evaluate the goal distribution of pedestrian trajectories to provide potential destinations for prediction. However, this process neglects the global perspective, weakening the global dependency among trajectory points, which can increase the divergence in the goal distribution and reduce prediction accuracy. To address this, we propose a Markov Random Field (MRF) clique model method for pedestrian trajectory prediction network. In this approach, each future trajectory point is treated as a node variable, forming cliques with historical trajectory points, thereby constructing a global clique model between them. Subsequently, clique models are established in both the goal network and the prediction network. This clique-based approach enhances the ability to capture and represent complex dependency in trajectories. Additionally, to further deepen feature fusion in the prediction network, we design a fusion network based on the Kolmogorov-Arnold Network (KAN). Specifically, we propose MCK-net, which primarily includes a two-tiered clique goal module and a clique prediction module based on CK-LSTM (Clique and KAN LSTM). These designs achieve dual performance improvements in both the goal and prediction networks. Comparative experiments with various previous outstanding algorithms demonstrate that our MCK-net achieves state-of-the-art (SOTA) performance on all the scenes of the ETH/UCY datasets. Code will be available at: https://github.com/SunshineDHW/MCK-net