Introduction
In the field of autonomous driving, identifying a vehicle traveling at high speeds is a well-solved problem. However, detecting "subtle motions"—such as a car just beginning to edge out of a parking space or a vehicle starting to move from a stop sign—remains a significant challenge. These tiny movements are often the first indicators of a change in traffic dynamics, yet they are frequently missed by traditional LiDAR-based perception systems.
The Challenge: The "Swimming Effect"
The primary obstacle in detecting subtle motion is the "swimming effect." LiDAR sensors generate sparse point clouds that change from frame to frame as the sensor moves. This creates a phenomenon where static objects appear to shift or "swim" slightly. While this noise is negligible for high-velocity objects, it often overpowers the actual signal of subtle motions, leading to high rates of false positives and false negatives.
The S’More Framework
To address this, researchers developed S’More (Subtle Motion Regressor), an object-centric framework designed to isolate real motion from sensor noise. The process begins by voxelizing five consecutive LiDAR frames into a 3D occupancy grid.
The core innovation of S’More is Local Occupancy Completion. Instead of relying on sparse, flickering points, the model learns to "fill in" the visible surface of an object. By densifying the point cloud, the system creates a stable geometric anchor, making actual physical displacement much easier to distinguish from the swimming effect.
Following densification, an encoder-decoder network processes the grid through two specialized heads: a Motion Detector to classify the object as static or moving, and a Motion Flow Predictor to calculate the specific velocity vector for every occupied voxel.
Experimental Results
The researchers established a new Subtle Motion Benchmark using the Waymo Open Dataset, focusing on sequences where objects moved less than 0.2 meters over a 0.5-second window.
S’More demonstrated a significant performance leap over standard methods. It achieved an End-Point Error (EPE) of 0.0437m, nearly halving the error of the tracking-based CenterPoint baseline (0.0927m). Furthermore, its F1 score of 0.8323 significantly outperformed both ICP and CenterPoint in the subtle motion regime. Qualitative analysis confirmed that S’More effectively "rectifies" false positives by correctly identifying "swimming" parked cars as static through its understanding of completed geometry.
Conclusion and Future Work
S’More proves that densification is a powerful de-noising strategy for 3D perception. By treating the swimming effect as structured geometric noise rather than random jitter, the framework provides a robust solution for safety-critical "corner cases." Future work aims to extend this approach to non-rigid actors like pedestrians and cyclists, whose complex movements present additional layers of perceptual difficulty.
Summary updated