Abstract

Human motion prediction remains a central problem in computer vision, requiring models that simultaneously capture long-range temporal dependencies, preserve fine-grained spatial structure, and remain computationally efficient. This thesis introduces two complementary deep-learning frameworks, namely SimGate and GEAN, which share a common DCT-based embedding of joint trajectories and a gated multilayer perceptron (MLP) for spatial-frequency interaction, yet diverge in architectural philosophy. SimGate departs from the prevailing reliance on self-attention. By omitting attention layers it achieves linear computational complexity with respect to the number of input frames, while still encoding long-term context through a lightweight gated MLP that selectively modulates information between intra-frame spatial projections and inter-frame frequency projections, enabling effective spatio-temporal modeling without incurring quadratic overhead. A final prediction head maps the gated representation to future 3D joint positions. Experiments on Human3.6M, AMASS, and 3DPW demonstrate that SimGate attains competitive mean per-joint position error and perceptual motion quality, matching or surpassing state-of-the-art baselines, while maintaining a compact parameter budget and high inference speeds on a single GPU. GEAN builds upon the SimGate architecture by integrating efficient attention mechanisms to further enhance long-term memory dependency modeling. Quadratic scaling is alleviated through reformulated dot-product attention and key-dimensionality reshaping, thereby reducing overall parameter count. Attention outputs are interleaved with the gated MLP’s linear refinement, ensuring that only salient global information propagates toward the prediction head. As a result, GEAN consistently improves predictive accuracy over SimGate with only a modest increase in computational and memory cost, enabling robust modeling of complex motion patterns, including rapid transitions and abrupt directional changes. Taken together, the shared DCT-based representation and spatial-frequency gating establish a unifying theoretical framework across both models. Empirical results confirm that SimGate and GEAN generate temporally coherent motion sequences that closely adhere to ground-truth trajectories, preserving smooth joint evolution and anatomical plausibility even in the presence of occlusions or missing observations.