EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Spiking YOLO for Fire and Smoke Detection in Ground and UAV Images: A Comparison of Direct, Poisson, and Latency Spike Inputs

Pau Bujons, Raúl Parada

Fire · 06.11.2026 · 9 (11) · 495

DOI: 10.3390/fire9110495 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Fire, online erschienen am 06.11.2026. Verfasst von Pau Bujons und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Wildfire detection on camera-equipped UAVs and other edge platforms motivates the investigation of detectors suitable for power-constrained operation. This work investigates whether explicit RGB-to-spike encoding offers detection or computational advantages over direct RGB input in a spiking fire and smoke detector. We present a controlled comparison of direct RGB input, Poisson rate encoding, and latency encoding within SpikeYOLO, keeping the detector architecture, pretrained initialization, training data, and training recipe fixed across input configurations. The three representations are evaluated at T=4 and T=6 timesteps, jointly assessing detection performance, inference time, and GPU energy consumption, with YOLOv8m and a lightweight CNN as conventional references. Quantitative experiments use the filtered D-Fire dataset, complemented by qualitative experiments on UAV videos and an analysis of near-duplicates across dataset splits. On the 2786-image test set, direct-RGB, Poisson-encoded, and latency-encoded SpikeYOLO achieved mAP@0.5 scores of 0.701, 0.624, and 0.590, respectively, at T=4, compared with 0.753 for YOLOv8m. Under the evaluated training budget, increasing T to 6 provided no additional detection performance gains for direct RGB or Poisson encoding. GPU profiling showed similar network execution costs across input representations at a fixed timestep count and negligible encoding overhead, although network inference time and GPU energy consumption exceeded those of YOLOv8m. The near-duplicate analysis flagged 62% of test images as overlapping with training or validation scenes; evaluation of the remaining subset preserved the relative model performance at T=4. These findings characterize representation-dependent detection performance under a fixed training budget and show that sparse input encoding alone does not reduce execution cost in the evaluated dense GPU implementation, providing a reference for subsequent evaluation with event-based sensing and neuromorphic hardware.

Zitieren

APA

Bujons, P., & Parada, R. (2026). Spiking YOLO for Fire and Smoke Detection in Ground and UAV Images: A Comparison of Direct, Poisson, and Latency Spike Inputs. Fire, 9(11), 495. https://doi.org/10.3390/fire9110495

Vancouver

Bujons P, Parada R. Spiking YOLO for Fire and Smoke Detection in Ground and UAV Images: A Comparison of Direct, Poisson, and Latency Spike Inputs. Fire. 2026;9(11):495. doi:10.3390/fire9110495

BibTeX
@article{Bujons2026,
  author = {Bujons, Pau and Parada, Raúl},
  title = {Spiking YOLO for Fire and Smoke Detection in Ground and UAV Images: A Comparison of Direct, Poisson, and Latency Spike Inputs},
  journal = {Fire},
  year = {2026},
  volume = {9},
  number = {11},
  pages = {495},
  doi = {10.3390/fire9110495}
}
Kostenlos starten →