Abstract
In this paper, we propose an efficient event-based motion estimation framework for various motion models. Different from previous works, we design a progressive event-to-map alignment scheme and utilize the spatio-temporal correlations to align events. In detail, we progressively align sampled events in an event batch to the time-surface map and obtain the updated motion model by minimizing a novel time-surface loss. In addition, a dynamic batch size strategy is applied to adaptively adjust the batch size so that all events in the batch are consistent with the current motion model. Our framework has three advantages: a) the progressive scheme refines motion parameters iteratively, achieving accurate motion estimation; b) within one iteration, only a small portion of events are involved in optimization, which greatly reduces the total runtime; c) the dynamic batch size strategy ensures that the constant velocity assumption always holds. We conduct comprehensive experiments to evaluate our framework on challenging high-speed scenes with three motion models: rotational, homography, and 6-DOF models. Experimental results demonstrate that our framework achieves state-of-the-art estimation accuracy and efficiency.
Introduction
Event cameras, or bio-inspired silicon retinas, respond to pixel-wise brightness changes asynchronously, offering high temporal resolution and dynamic range. However, their high-speed nature generates millions of events per second, creating a significant computational bottleneck for real-time applications. Traditional motion estimation methods, such as contrast maximization or entropy minimization, typically process every event in a batch to find optimal camera motion. This approach is often redundant because events occurring close together in time carry nearly identical information. Furthermore, fixed batch sizes often fail when camera velocity changes rapidly, violating the "constant velocity" assumption required for accurate estimation.
Methodology
The proposed Progressive Spatio-temporal Alignment (PEME) framework introduces several innovations to overcome these challenges:
Time-Surface Map Alignment and Sampling
Instead of performing costly pairwise event comparisons, PEME aligns events to a Time-Surface (TS) map—a 2D grid storing the timestamps of the most recent events. Motion parameters are updated by minimizing a novel TS loss. A key insight of this work is that processing all data is unnecessary; sampling a tiny fraction of events (e.g., 1,000 out of a 30,000-event batch) provides sufficient gradient information to guide optimization, drastically reducing runtime.
Progressive Refinement
The framework iteratively refines motion parameters using a progressive scheme. By employing bidirectional warping (warping events both forward and backward in time), the system creates more accurate reference targets for alignment, effectively handling the spatial sparsity inherent in event data.
Dynamic Batch Strategy
To ensure the constant velocity assumption holds regardless of scene complexity or camera acceleration, the algorithm adaptively appends event bundles to a batch only if they are consistent with the current motion model, as determined by an overlap ratio.
Experimental Results
The framework was evaluated on major benchmarks including the Event-Camera Dataset, DVSMOTION20, and MVSEC, covering rotational, homography, and 6-DOF models.
- Efficiency: The "Ours-1k" variant achieved a processing time as low as 0.3 µs per event, making it significantly faster than previous state-of-the-art methods like CMax and ST-PPP.
- Accuracy: Despite the sparse sampling, the framework achieved superior accuracy. In the "shapes_rotation" dataset, it reduced mean angular velocity error by over 70% compared to existing real-time methods.
- Real-time 6-DOF: PEME is one of the few frameworks capable of real-time 6-DOF motion estimation on a standard CPU, maintaining high trajectory fidelity in challenging indoor flying sequences.
Conclusion
PEME demonstrates that efficient motion estimation can be achieved by leveraging spatio-temporal correlations and sparse sampling. By progressively aligning events to a time-surface map and utilizing a dynamic batching strategy, the framework achieves state-of-the-art accuracy and real-time performance. Future work will focus on integrating this framework into full SLAM systems for edge robotics platforms.
Summary updated