Accurate trajectory forecasting remains challenging due to insufficient feature representation and uncertainty modeling in existing methods. We propose a Diffusion-Based Cross-Modal Fusion Model (DCF) that enhances predictive performance through comprehensive spatiotemporal modeling and probabilistic refinement. Firstly, we develop polar coordinate-based spatiotemporal encoding that decomposes motion patterns into radial velocity and angular components. This representation effectively captures both short-term collision avoidance behaviors and long-term movement trends while maintaining rotation invariance across different scales. Secondly, we design a cross-modal attention fusion module that integrates four complementary feature streams: sequential trajectory patterns, visual motion features from autocorrelation decoder, periodic motion features, and skeletal keypoint dynamics. Feature enhancement techniques are applied to preserve modality-specific information while learning cross-modal interactions. Finally, we implement a diffusion-based uncertainty refinement framework that progressively denoises trajectory distributions through reverse diffusion processes. This mechanism adaptively models various uncertainty sources including individual motion randomness and environmental perturbations, generating physically plausible future trajectories. Extensive evaluations on ETH/UCY, Stanford Drone Dataset, and ActEV/VIRAT benchmarks demonstrate our model’s superior performance.