4DRecon: Spatio-Temporal Reconstruction of Multiple Dynamic Objects in LiDAR Streams

Dynamic 4D reconstruction (3D + time) of multiple rigid objects from sequential LiDAR scans presents critical challenges for autonomous driving. While existing learning-based methods often struggle to generalize to unseen scenarios, traditional model-based approaches typically demonstrate computational inefficiency when processing multiple objects. This paper introduces 4DRecon, a novel model-based framework that departs from conventional detection-tracking pipelines by integrating three core functionalities: (1) efficient extraction of dynamic points directly from raw LiDAR streams, enabling subsequent object detection; (2) simultaneous object tracking and shape reconstruction using an Error-State Iterative Kalman Filter (ESIKF) that iteratively refines both pose estimation and geometric representation; and (3) zero dependency on training data, which fundamentally enables cross-platform generalizability under any sensor or geographic variation. For comprehensive evaluation, we establish rigorous testing protocols utilizing both the KITTI dataset and a newly annotated benchmark comprising 1,121 high-quality tracklets from the Waymo dataset. Experimental results demonstrate that our approach achieves a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$25 \times$</tex> speed improvement (0.01113 s vs. 0.2797 s per object) compared to state-of-the-art model-based methods while maintaining consistent reconstruction accuracy (less than 20 cm error) even in challenging low-point-density scenarios. The framework's purely architecture effectively overcomes the generalization limitations inherent in data-driven methods.