Gate-Calibrated Double Disentangled Distribution Matching Network for Cross-Domain Pedestrian Trajectory Prediction

In cross-domain pedestrian trajectory prediction, most existing methods usually focus on learning entangled spatial-temporal domain-invariant features, while ignoring the different contributions of spatial and temporal shifts to the prediction model. To address this issue, we propose a novel gate-Calibrated Double Disentangled Distribution Matching Network (CD<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{3}$</tex-math></inline-formula>MN) that can effectively eliminate cross-trajectory domain shifts at both the spatial and temporal levels while learning robust prediction using a calibrated gated-fusion. The key idea of CD<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{3}$</tex-math></inline-formula>MN is to model domain-invariant features across trajectories as a calibrated gated-fusion of disentangled domain-invariant features at the temporal and spatial levels. We first introduce a spatial-temporal disentanglement module to disentangle the spatial-temporal properties of pedestrian trajectories from the spatial-level and temporal-level. Secondly, we design a domain-invariant disentanglement module for learning domain-invariant sample-level transferable feature representations at the spatial and temporal levels. Finally, to effectively fuse these disentangled temporal and spatial features, we design a calibrated gated-fusion module where both inter-level and intra-level knowledge are introduced to calibrate the fusion gate. Extensive experiments on real datasets demonstrate the effectiveness of CD<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{3}$</tex-math></inline-formula>MN.