Abstract: Achieving camera pose estimation and 4D Gaussian radiance field reconstruction in dynamic scenes serves as a critical pathway for bridging 2D image observations and real-world spatiotemporal modeling.Unlike existing methods that typically treat dynamic objects as noise to be removed,this paper proposes a motion-aware 4D reconstruction framework tailored for dynamic environments.The system incrementally tracks camera poses and constructs a time-varying Gaussian radiance field from continuous RGB-D image sequences.Specifically,motion masks are generated to provide per-pixel static or dynamic priors; 3D Gaussian primitives are then divided into static and dynamic subsets,where dynamic Gaussians are modeled via a sparse set of control points and an MLP-based deformation network to capture non-rigid temporal motions.To further enhance inter-frame motion modeling,we introduce a novel Gaussian-rendered optical flow reconstruction approach that explicitly renders motion offsets of dynamic objects between adjacent frames.These optical flow cues are integrated with conventional photometric and geometric consistency constraints to jointly optimize the 4D Gaussian field.Experimental results on the TUM RGB-D and BONN dynamic datasets demonstrate the proposed method achieves robust pose tracking and high-quality scene reconstruction,especially in complex scenes with multiple moving objects and structural variation.

Key words: dynamic scenes, 4D Gaussian radiance field, pose estimation, motion awareness, sparse control points, optical flow supervision, joint optimization