Arxiv 2502.18490 surveys event cameras in human-centered applications — lessons that apply beyond the lab

Most event-camera research we track is explicitly defense or aerospace adjacent — drone detection, guidance, space surveillance. But a large body of work on event cameras comes from a different direction: human-computer interaction, rehabilitation, ergonomics, sports science. Applications like gesture recognition, eye tracking, and action analysis drove significant event-camera hardware development and produced algorithm families that transfer well to other domains.

Arxiv 2502.18490 surveys this body of work systematically. Reading it from a terminal guidance perspective reveals several algorithm patterns worth adapting.

Gesture recognition and the sparse detection analogy

Hand gesture recognition from event cameras is a prototypical fast-detection task: the hand is a small, fast-moving object against a relatively static background, and the system needs to classify its motion trajectory in real time. The detection and classification pipelines developed for this application — event clustering, trajectory association, lightweight classification networks — have direct structural parallels to the C-UAS detection pipeline.

The survey catalogs which network architectures perform best for sparse, asynchronous event streams in the gesture domain. Graph neural networks appear repeatedly as the most data-efficient approach for event-based classification tasks where the spatial extent of the object is limited and the temporal pattern of events carries most of the information. This is consistent with what the ACGR and DAGr papers we’ve covered show for object detection — the graph approach handles irregular data naturally and avoids the wasted computation of dense-grid convolutional processing.

Gaze tracking and centroid stability

Eye tracking from event cameras is a precision centroid estimation problem: the pupil is a small, dark region moving against a less-dark sclera, and you need to know its center to within a fraction of a degree. At fast saccade speeds — 700 degrees per second for a reflex saccade — frame cameras produce unacceptable blur. Event cameras handle the saccade speed without blur, and the per-pixel logarithmic response gives good contrast between the pupil and sclera independent of illumination level.

The centroid estimation algorithms developed for gaze tracking are directly applicable to target tracking in a seeker. The pupil centroid problem and the drone target centroid problem differ in the background (a structured iris versus a sky background) but are the same estimation problem structurally. Pipeline components from gaze tracking research — adaptive thresholding on the event stream, temporal filtering of the centroid estimate, handling of partial occlusion — are worth surveying for the seeker application.

The power constraint

Several papers in the survey target low-power deployment on wearable and embedded devices — smart glasses, EEG headsets, medical monitoring equipment. This is a useful technology pressure: the algorithm and hardware choices that work at 10–100 mW on a wearable device overlap substantially with what’s possible on a small-form-factor airborne processor. The survey identifies specific network compression and quantization approaches that achieve near-baseline accuracy at 4-bit precision on fixed-point hardware, which is directly relevant for FPGA deployment.

What the human-centered domain contributes

The practical contribution of the human-centered event-camera literature to our work is threefold. First, it provides a large body of labeled event-camera data — gesture datasets, gaze datasets, action recognition datasets — that can serve as pre-training material for networks later fine-tuned on defense-relevant target data. Second, it provides a literature on real-time event-camera processing under tight compute constraints, because wearable applications have power budgets stricter than many airborne applications. Third, it provides end-to-end validated pipelines that have shipped in consumer products — eye tracking is now in high-end gaming monitors and VR headsets — which means some of the processing infrastructure has been production-proven, even if not in the relevant application domain.

The transfer isn’t direct — gesture recognition doesn’t encounter muzzle blast disturbance, and eye tracking doesn’t have to contend with clutter rejection at 200 meters — but the underlying sensor characterization and algorithm families are shared, and ignoring the human-centered literature means ignoring a large fraction of the relevant engineering work.