GeoIMO: Geometry-Driven Independent Motion Classification for Event Cameras

Autonomous driving systems require the ability to distinguish between stationary scene elements and independently moving objects (IMOs) to ensure safety. Existing automotive datasets often rely on frame-based annotations that fail to capture the motion-specific characteristics of event cameras. This paper addresses the need for a lightweight, annotation-free framework that can classify objects as static or moving by exploiting the geometric structure of ego-motion latent within event streams.

The proposed framework, GeoIMO, operates without the need for manual labels or learned components. It estimates global background ego-motion using a Focus of Expansion (FOE) model, which is augmented with yaw compensation to account for vehicle steering. By comparing the motion of events within a given bounding box to this global ego-motion estimate, the system calculates a scale-invariant residual. Objects are classified as independently moving if their local motion significantly deviates from the expected background motion. The framework also incorporates temporal stabilization to improve robustness across consecutive event windows.

Experiments on the MVSEC and Prophesee 1 Megapixel Automotive Detection datasets demonstrate that GeoIMO provides consistent motion classification across diverse driving scenarios. The inclusion of yaw compensation is critical for maintaining accuracy during turns, where pure translational models fail. While a radial local motion model is theoretically more expressive, the authors find that a simpler translational model offers a more favorable accuracy-efficiency trade-off, as the performance gains from the radial model are marginal given the sparsity of event data within bounding boxes.