Event cameras excel at detecting small, fast drones, but today's detectors give away their key advantage: they convert the sparse event stream into dense grids and pay dense-processing cost on inputs that are almost entirely empty. We present SparseVoxelDet, to our knowledge the first coordinate-sparse 3D event voxel bounding-box detector: backbone, feature pyramid, temporal reduction, and detection head all operate directly on coordinate-indexed features, with no dense spatial grid at any stage. Building it exposed a hidden failure mode we name support inflation: an input filling a median 0.0652% of the voxel lattice inflates stage by stage until standard pyramid fusion leaves the detection head locally near-dense. We answer with two ideas. Expansion-free inverse-convolution fusion provably creates no active sites beyond the stored backbone supports, cutting head occupancy from a median 78.88% to 10.53%; quality-aligned supervision then recovers more accuracy than preserving sparsity costs. The payoff is measured, not assumed: executing the same trained network densely costs a median 27.5x the work and 4.65x the latency across 5,000 paired frames, with no frame cheaper dense, while the 6.22M-parameter detector reaches 87.01 AP50 on the FRED drone benchmark, ahead of matched dense controls, and holds its lead on the held-out test partition evaluated once. Sparsity, preserved by construction and supervised well, delivers the efficiency and the accuracy together.
Figures & tables
Figure 1: Dense vs. coordinate-sparse processing. Left: a dense detector processes 409,600 pixels at 6402 . Right: SparseVoxelDet input has a median 4,272 active coordinates, 95.88 × fewer than dense input pixels on a median-count basis. Under transpose fusion, pyramid expansion and temporal pooling grow support to a median 20,194.5 of 25,600 stride-4 sites (78.88%); the expansion-free fusion of Section 3.3 returns the detection head to a median 10.53% occupancy, with the full distribution and its tail reported in Section 5.4 . These full-development-split, stage-qualified counts describe topology, not FLOPs or memory.
Method
Sparsity mechanism
B
N
H
Dense spatial grid enters
RVT ( Gehrig and Scaramuzza, 2023 )
– (dense baseline)
×
×
×
input
SAST ( Peng et al., 2024 )
token/window masking
×
×
×
input
SMamba ( Yang et al., 2025 )
token dropping
×
×
×
input
SFOD ( Fan et al., 2024 )
activation (spiking)
×
×
×
input
EAS-SNN ( Wang et al., 2024 )
activation (spiking)
×
×
×
input
SCR ( Wang et al., 2025 )
activation (thresholded)
×
×
×
input
Table 1: Where dense spatial grids enter event-based detectors (and the closest LiDAR analogue). B/N/H: backbone / neck / head operate on structurally sparse coordinate-indexed features. “Input” means a dense spatial grid is constructed from the raw events before learned computation. † Dense output verified in the official implementations: sparse features are converted to a dense map before a fully-connected YOLO layer (Messikommer et al.), or graph nodes are max-pooled onto a fixed grid before a linear YOLO output (AEGNN). ∗ Sparse concatenation of downsampled stages; no multi-scale pyramid.
Figure 2: SparseVoxelDet architecture. Events are binned into a sparse 3D voxel grid ( T×H×W , 6 temporal-surface channels). The SparseSEResNet backbone processes voxels through four stages of sparse 3D residual blocks with SE attention at stages 3–4, producing multi-scale features { C2 , C3 , C4 } at strides 4, 8, 16. The expansion-free sparse pyramid fuses scales top-down via sparse lateral projections and inverse convolutions that restore coarse features onto the stored backbone supports, yielding a unified stride-4 feature map without creating new active sites. After temporal max-pooling collapses the time axis, the detection head (shared MLP) predicts classification, LTRB box regression, and box quality at each active 2D position, followed by NMS. The spatial feature path constructs no dense H×W or T×H×W feature grid.
Figure 3: FRED dataset samples across lighting conditions. Top row: RGB frames. Bottom row: the dataset’s own event-frame renderings, in which events appear as light blue and white marks against a dark background; the corner inset enlarges the target region, with the ground-truth box drawn inside it. While RGB quality degrades from day to night, the event stream retains a visible drone signature in each. Qualitative illustration; no per-condition accuracy is claimed.
Active sites
Stage sites
Occupancy (%)
Stage
(median)
(reference)
median
p95
max
Input volume
4,272
6,553,600
0.0652
0.77
5.73
Stem
14,266
1,638,400
0.87
11.91
33.82
Backbone C2 (stride 4)
19,271
409,600
4.70
53.33
80.72
Backbone C3 (stride 8)
16,867
102,400
16.47
92.92
99.86
Backbone C4 (stride 16)
11,742
25,600
45.87
99.94
100
Table 2: Stage-qualified SparseVoxelDet topology at 6402 , measured over all 103,672 frames of the 37-sequence development-validation split. Occupancy is the realized active-site share of each stage’s own T × H × W site count; these counts are not interpreted as FLOPs or memory. Read down the occupancy column: the median rises at every stage from the input to the fused pyramid, which is support inflation measured rather than asserted. The final row is the expansion-free head, which replaces the row above it and is not a further step in that ladder. The input carries a second, coarser basis used in the text (its median 4,272 active coordinates are ∼ 95.88 × fewer than the 409,600 pixels of the 6402 lattice, at a mean of 14,582.66), and dividing those two numbers returns 1.04%, which is a coordinate count against a pixel lattice and not an occupancy.
Figure 4: Event preprocessing and comparison paths. Both arms use non-overlapping 33 ms event windows of the same nominal duration, paired to labels under slightly different indexing conventions (Appendix 9 ). The primary SparseVoxelDet path anisotropically maps native 1280×720 coordinates onto a 640×640 lattice, quantizes timestamps into T=16 bins, suppresses voxels whose count exceeds max(3,3cˉ) for that window, computes six temporal-surface features, and retains only coordinate–feature pairs. The pipeline also supports a native-resolution sparse configuration, whose voxelized data we release, but no result reported in this paper is measured at it. The dense event-only control renders the same event stream at native aspect ratio and letterboxes it to 640×640 , yielding 640×360 signal plus 280 padding rows. Section 6 accounts for the input-sampling asymmetry between the two arms.
Model
Processing
Modality
AP50
AP50:95
Development-validation split (103,672 frames)
YOLO11n control (letterbox)
Dense
Event
77.29
38.12
YOLO11n control (resolution-matched)
Dense
Event
84.68
42.44
SparseVoxelDet
Sparse
Event
87.01
43.14
Paired margin vs. letterbox, pooled
+ 9.72
+ 5.02
per-sequence min
− 12.78
− 30.48
Table 3: Controlled comparison on FRED, in two blocks: development-validation estimates (all rows on the same frames, labels, evaluator, and post-processing) and the canonical test partition evaluated once, two seeds as two values. The YOLO11n rows are a system comparator, not an architecture-matched baseline; Section 6 states the asymmetries, Section 5.1 the dispersion protocol. The resolution-matched control is a single seed on each arm and has no sealed row: the single permitted test evaluation was spent before it was scored. Per-sequence rows re-score the same frozen caches sequence by sequence; average precision is not additive over sequences, so the pooled margin is not their mean.
Fusion
Supervision
AP50
AP50:95
Head occ. (%)
Transpose
Center-only
86.31
40.54
78.88
Inverse conv.
Center-only
85.71
40.23
10.53
Inverse conv.
Center-only (topology-matched)
85.52
40.25
10.53
Inverse conv.
Quality-aligned
87.01
43.14
10.53
Table 4: Supervision and fusion ablation on the development protocol. All rows share the SparseVoxelDet backbone, data, nominal 20-epoch schedule, evaluator, and selection rule; the topology-matched center-only row shares the quality-aligned row’s exact 3-GPU distributed configuration, while the other center-only rows trained on one accumulating GPU (Section 6 ). Head occupancy is the median realized share of stride-4 post-pool sites over the full development-validation split. The table holds one factor fixed at a time rather than crossing them: with supervision fixed at the center-only objective, the fusion swap cuts head occupancy from 78.88% to 10.53% and costs 0.60 AP50; with fusion fixed at the expansion-free architecture, quality-aligned supervision returns 1.49 AP50 at matched topology. The untrained factorial cell (quality-aligned supervision on transpose fusion) leaves the two changes not separated factorially; the final configuration ties for the sparsest head occupancy and is the most accurate of the four trained arms.
Figure 5: Development-validation trajectories of the two supervision arms across the 19 epochs both arms scored, under the frozen evaluator. The quality-aligned arm leads at 18 of those 19 epochs on AP50 and on AP50:95; the single exception is epoch 0, before either arm has converged. The star marks the checkpoint the frozen selection rule returns: epoch 5 of the quality-aligned arm, released as epoch_005.pt , and the dotted line carries it into the AP50:95 panel. The arms differ in optimizer topology as well as in supervision, so the comparison is matched by epoch and not by update count (Section 6 ).
Figure 6: Error forensics for the selected checkpoint on the 37-sequence development-validation split, under the frozen evaluator at score 0.05, class-agnostic NMS 0.5, and at most 100 detections per frame. (a) Every unmatched ground-truth box at matching IoU 0.50 and τ∗=0.25 is assigned exactly one cause under a strict priority order, so the five causes sum to the false-negative total exactly; the axis is logarithmic and the complete-miss bar is highlighted. (b) Confidence histogram of true- and false-positive detections over the full scored set, with τ∗ marked. The two populations cross once, low on the confidence axis: above the crossing the detector is right far more often than wrong, and the mass that costs recall lies below it.
Figure 7: Matching-threshold sensitivity of the selected checkpoint on the development-validation split under the frozen evaluator. (a) Average precision across matching IoU thresholds from 0.50 to 0.95. (b) Ground-truth boxes recovered at the reported operating point τ∗=0.25 , against the boxes the scored detection set already reaches before the confidence cut is applied; the shaded band is the share that the confidence threshold alone governs. Both panels are computed from the frozen offline prediction cache that scores the selected checkpoint at 87.01 AP50, the same path as Table 3 and Table 4 . The dashed line is the candidate-coverage ceiling: 99.75% of ground-truth boxes are overlapped by at least one candidate above the 0.001 pre-NMS score floor, so the zero-overlap cohort that no re-ranking and no threshold change can reach is 302 boxes of 121,982.
Figure 8: Matched detections from the selected checkpoint on the development-validation split, one per quartile of ground-truth target size, arranged in two rows of two. Each cell pairs the accumulated event frame (top) with the RGB frame at the same annotation index (bottom); green: ground truth, cyan: prediction. Annotations give the sequence, frame, drone model, matched IoU, detection score, and the number of active pixels inside the ground-truth box. Cells were selected by the size-quartile rule of Section 5.5 , not chosen for appearance; they illustrate the operating point and are not a sample designed to represent the error distribution.
Figure 9: The four principal false-negative causes of Figure 6 , one instance each, from the selected checkpoint on the development-validation split. Green: ground truth; orange: the best-overlapping candidate where one exists. Titles give each cause’s share of all 15,215 false negatives. The below-confidence cell draws one of the 6,335 instances whose correct box survives non-maximum suppression, so the box shown is itself the sub-threshold detection; the remaining 3,950 of that group have no surviving box to draw. Every number printed on a cell is recomputed from the box drawn in that cell. The complete-miss cell contains no active pixel inside the ground-truth box, the cohort-wide property reported in Section 5.5 .
Event cameras offer high temporal resolution and motion sensitivity for tiny UAV detection, yet distant targets generate sparse and fragmented events that are easily overwhelmed by clutter and ego-motion. Existing methods mainly rely on dense event representations or local sparse spatiotemporal modeling, resulting in redundant computation or fragmented modeling of motion continuity across distant asynchronous events. To address this limitation, we introduce serialized motion evidence accumulation, which treats motion continuity as an ordered evidence propagation process. Specifically, the same event stream is organized into locality-preserving spatiotemporal paths and chronology-preserving temporal paths through the latent complementary serializations. Based on this principle, we propose PointEvent, a lightweight event-wise state-space framework that alternates serialized scans across the complementary orders, progressively consolidating fragmented motion evidence beyond fixed local neighborhoods. A high-resolution event branch preserves fine-grained target responses, while compact context modulation suppresses interference. Experiments demonstrate that PointEvent achieves SOTA with the fewest parameters and fastest measured inference among the compared methods. Code: https://github.com/wzz-z/PointEvent
Zongze Wu, Baofeng Jia, Weiqi Yan +4
State key Lab of Extreme Environment Optoelectronic Dynamic Testing Technology and Instrument, Nanjing University of Science and Technology, China · Jiangsu Key Lab of Visual Sensing and Intelligent Perception, Nanjing University of Science and Technology, China · Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University, China
The rapid proliferation of drones requires balancing innovation with regulation. To address security and privacy concerns, techniques for drone detection have attracted significant attention.Passive solutions, such as frame camera-based systems, offer versatility and energy efficiency under typical conditions but are fundamentally constrained by their operational principles in scenarios involving fast-moving targets or adverse illumination.Inspired by biological vision, event cameras asynchronously detect per-pixel brightness changes, offering high dynamic range and microsecond-level responsiveness that make them uniquely suited for drone detection in conditions beyond the reach of conventional frame-based cameras.However, the design of most existing event-based solutions assumes a static camera, greatly limiting their applicability to moving carriers--such as quadrupedal robots or unmanned ground vehicles--during field operations.In this paper, we introduce a real-time drone detection system designed for deployment on moving carriers. The system utilizes a spinning event-based camera, providing a 360° horizontal field of view and enabling bearing estimation of detected drones. A key contribution is a novel image-like event representation that operates without motion compensation, coupled with a lightweight neural network architecture for efficient spatiotemporal learning. Implemented on an onboard Jetson Orin NX, the system can operate in real time. Outdoor experimental results validate reliable detection with a mean angular error below 2° under challenging conditions, underscoring its suitability for real-world surveillance applications. We will open-source our complete pipeline to support future research.
Kuan Dai, Hongxin Zhang, Sheng Zhong +1
Neuromorphic Automation and Intelligence Lab (NAIL) at School of Artificial Intelligence and Robotics, Hunan University, Changsha, China.
Detecting fast-moving objects, such as unmanned aerial vehicle (UAV), from event camera data is challenging due to the sparse, asynchronous nature of the input. Traditional Discrete Fourier Transforms (DFT) are effective at identifying periodic signals, such as spinning rotors, but they assume uniformly sampled data, which event cameras do not provide. We propose a novel per-pixel temporal analysis framework using the Non-uniform Discrete Fourier Transform (NDFT), which we call Drone Detection via Harmonic Fingerprinting (DDHF). Our method uses purely analytical techniques that identify the frequency signature of drone rotors, as characterized by frequency combs in their power spectra, enabling a tunable and generalizable algorithm that achieves accurate real-time localization of UAV. We compare against a YOLO detector under equivalent conditions, demonstrating improvement in accuracy and latency across a difficult array of drone speeds, distances, and scenarios. DDHF achieves an average localization F1 score of 90.89% and average latency of 2.39ms per frame, while YOLO achieves an F1 score of 66.74% and requires 12.40ms per frame. Through utilization of purely analytic techniques, DDHF is quickly tuned on small data, easily interpretable, and achieves competitive accuracies and latencies to deep learning alternatives.