ASUMOT: Motion-Consistency-Based Asynchronous UAV Detection and Tracking with Event Cameras

Low-altitude UAV surveillance using event cameras is often hindered by the sparse and fragmented nature of event responses, especially for long-range targets. Conventional asynchronous trackers often treat every detected event blob as an independent object, leading to duplicate trajectories and unstable identities. This paper asks: can we improve UAV tracking by decoupling low-level event blob detection from high-level target identity maintenance?

The authors propose ASUMOT, a framework that operates directly on raw event streams. Instead of assuming a one-to-one mapping between an event blob and a physical UAV, ASUMOT represents each UAV as a time-varying collection of motion-consistent blobs. The pipeline consists of three main stages:

1. Local Motion-Consistency Estimation: A trigger mechanism that uses the temporal orientation of local event structures to identify reliable candidates while suppressing background noise.
2. Multi-Task Verification: A lightweight neural network that evaluates candidate blobs to predict UAV confidence and motion direction, providing robust initialization.
3. Motion-Consistency Clustering (MCC): An association mechanism that aggregates fragmented, motion-consistent blobs into unified, identity-consistent UAV tracks.

Additionally, the authors introduce ES-UAV, a high-definition event-level benchmark with dense semantic annotations, designed to evaluate tracking performance in complex, cluttered urban environments.

ASUMOT demonstrates a superior accuracy-efficiency trade-off compared to existing asynchronous tracking methods. By aggregating fragmented blobs, the framework significantly reduces duplicate trajectories and identity switches. The proposed ES-UAV benchmark provides a rigorous testbed for evaluating these capabilities, showing that ASUMOT maintains stable tracking even when targets are small (under 32x32 pixels) and subject to complex maneuvers. The system is designed for edge deployment, maintaining low latency by processing events asynchronously.