A 360-degree Multi-camera System for Blue Emergency Light Detection Using Color Attention RT-DETR and the ABLDataset

Francisco Vacalebri-Lloret; Lucas Banchero; Jose J. Lopez; Jose M. Mossi

cs.CVMar 5, 2026

A 360-degree Multi-camera System for Blue Emergency Light Detection Using Color Attention RT-DETR and the ABLDataset

Authors: Francisco Vacalebri-Lloret, Lucas Banchero, Jose J. Lopez, Jose M. Mossi

Organizations: Universitat Politècnica de València, Spain · Institute of Telecommunications and Multimedia Applications, Universitat Politècnica de València, 46022 València, Spain.

Abstract

This study presents an advanced system for detecting blue lights on emergency vehicles, developed using ABLDataset, a curated dataset that includes images of European emergency vehicles under various climatic and geographic conditions. The system employs a configuration of four fisheye cameras, each with a 180-degree horizontal field of view, mounted on the sides of the vehicle. A calibration process enables the azimuthal localization of the detections. Additionally, a comparative analysis of major deep neural network algorithms was conducted, including YOLO (v5, v8, and v10), RetinaNet, Faster R-CNN, and RT-DETR. RT-DETR was selected as the base model and enhanced through the incorporation of a color attention block, achieving an accuracy of 94.7 percent and a recall of 94.1 percent on the test set, with field test detections reaching up to 70 meters. Furthermore, the system estimates the approach angle of the emergency vehicle relative to the center of the car using geometric transformations. Designed for integration into a multimodal system that combines visual and acoustic data, this system has demonstrated high efficiency, offering a promising approach to enhancing Advanced Driver Assistance Systems (ADAS) and road safety.

Figures & tables

Fig. 1: Illustrative example of the proposed system.

Fig. 2: Sample of the variety of emergency vehicles considered in ABLDataset.

Fig. 3: Levels of blue light grouping in emergency vehicles.

Fig. 4: Examples of labeled images from the ABLDataset.

Fig. 5: Flowchart of the dataset creation process.

Fig. 6: Explanatory diagram of the lens distortion procedure for data augmentation.

Model	P(%)	R(%)	mAP 0.5	mAP 0.5:0.95	$t_{i}$ (ms)
Retinanet	61.3	69.3	57.8	31.0	5.3
Faster R-CNN	77.3	75.2	80.1	49.2	5.7
Yolov5 – M	81.9	75.6	82.5	51.8	2.0
Yolov8 – M	85.8	72.7	82.4	50.9	2.2
Yolov10 – M	83.2	70.8	80.6	50.5	2.4
RT-DETR	85.0	82.8	89.1	53.2	3.7

TABLE I: Preliminary comparison of standard metrics for different object detection models (no modifications). These results guide the selection of the baseline model.

Fig. 7: Example of overlapping detections on the same ground truth.

Fig. 8: Example of vehicle-based annotation. The image shows a scene where four emergency vehicles are present, each labeled as a distinct vehicle in the dataset with its corresponding bounding box annotations.

Fig. 9: Color Attention - RT-DETR Diagram.

Fig. 10: Multi-Camera Fisheye System Configuration for 360° Coverage.

Fig. 11: Example of a dashboard alert design for the driver, indicating the presence of an emergency vehicle.

Model	P(%)	R(%)	mAP 0.5	mAP 0.5:0.95	$t_{i}$ (ms)
RT-DETR	85.0	82.8	89.1	53.2	3.7
CA RT-DETR (Ours)	88.8	86.4	91.6	56.1	4.1

TABLE II: Standard (non-contextualized) metrics for the customized model.

Fig. 12: Comparison of metrics in Bulb Array grouping for different confidence values.

Fig. 13: Comparison of metrics in Vehicle grouping for different confidence values.

Fig. 14: RT-DETR vs. CA RT-DETR vehicle metrics. The figure shows that incorporating the Color Attention module enhances vehicle detection performance by improving both precision and recall.

	Bulb Array			Vehicle
Model	Best $F1_{BA}$	$P_{BA}$	$R_{BA}$	Best $F1_{V}$	$P_{V}$	$R_{V}$
Retinanet	65.7	63.9	67.6	49.4	50.0	76.0
Faster R-CNN	81.39	84.0	78.9	77.93	70.1	87.7
Yolo v5 - MU	82.2	95.7	72.0	88.9	92.8	85.4
Yolo v8 - M	82.5	95.5	72.6	89.4	91.3	87.6
Yolo v10 - M	81.4	96.0	70.6	89.4	92.4	86.6

TABLE III: Best F1 Scores with Corresponding Precision and Recall for Bulb Array and Vehicle Detection.

Panel of Figure 18 — Fig. 15: Confusion matrices for the RT-DETR ( 15a ) and CA RT-DETR (ours) ( 15b ) models. The CA RT-DETR model outperforms the baseline, showing a 1.9% increase in true positives, a 35.2% decrease in false negatives, and a 23.1% reduction in false positives.

Fig. 16: Field test images capturing both static and dynamic scenarios, using instrumented vehicles to measure and simulate real-world emergency vehicle presence.

	Day			Night
Distance (m)	Det.	DR (%)	$\mathbf{E_{\varphi}(^{\circ})}$	Det.	DR (%)	$\mathbf{E_{\varphi}(^{\circ})}$
10	✓	100.0	$\sim 1.2$	✓	100.0	$\sim 1.1$
20	✓	100.0	$\sim 1.9$	✓	100.0	$\sim 1.2$
30	✓	98.3	$\sim 2.0$	✓	100.0	$\sim 2.4$
40	✓	63.4	$\sim 2.2$	✓	99.2	$\sim 2.5$
50	✓	22.9	$\sim 2.2$	✓	93.1	$\sim 2.6$

TABLE IV: System performance in static tests at different distances, detailing whether detection was achieved (Det.), the detection rate (DR), and the average azimuthal (

\mathbf{E_{\varphi}(^{\circ})}

) under both daytime and nighttime conditions.

Fig. 17: System detection ranges in static tests under daytime and nighttime conditions, illustrating how far the system can reliably detect active emergency lights.

		Day		Night
Scenario	Approx Vel. (km/h)	DR (%)	Dist. (m)	DR (%)	Dist. (m)
Frontal approach	30	92.5	2-50	99.8	2-50
Rear approach	30	91.2	2-50	99.7	2-50
Lateral approach	50	93.7	2-50	99.9	2-50
360º loop around car	20	99.6	20-30	100.0	20-30
Semicircular route (180º)	40	90.2	50-1.5	99.2	50-1.5

TABLE V: System performance in dynamic tests simulating various emergency vehicle approaches. The table reports the approximate speed, detection rate as a percentage, and the distance range between the measurement vehicle and the emergency vehicle in which the test was conducted, under both daytime and nighttime conditions.

Francisco Vacalebri-Lloret was born in 1998 in La Vila Joiosa, Alacant, Spain, he earned his Bachelor’s degree in Sound and Image Engineering in Telecommunications. He subsequently completed a Master’s degree in Telecommunications Engineering as well as a Master’s degree in Electronic Systems. With experience in the robotics industry, his research is focused on computer vision and artificial intelligence applied to visual systems and multimodal analysis. His work emphasizes the development and integration of advanced algorithms, sensor fusion techniques, and robust system architectures to address complex challenges in autonomous and intelligent systems.

Table 23

Lucas Banchero was born in Uruguay. He received his degree in Telecommunications Engineering with a specialization in Sound and Image from the University of Málaga, where he also completed a master’s degree in Acoustic Engineering. Since 2023, he has been pursuing a Ph.D. in Telecommunications at the Polytechnic University of Valencia. His research focuses on the industrial applications of sound, leveraging artificial intelligence for tasks such as acoustic scene analysis, signal processing, and sound-based event detection. His work spans various sectors, including the automotive industry, where he contributes to the development of Advanced Driver Assistance Systems (ADAS), and heavy industry, where he explores intelligent sound protection systems to enhance worker safety.

Table 24

Jose J. Lopez (Member, IEEE) was born in Valencia, Spain in 1969. He received M.S. and Ph.D. degrees in Telecommunications Engineering from Universitat Politècnica de València (UPV) in 1992 and 1999, respectively. He joined the teaching career in 1993 as assistant lecturer and has progressed through various teaching positions until becoming Full Professor in February 2011, a position he continues to hold. Dr. Lopez has supervised 10 doctoral theses and has served as Deputy Director of Innovation and Technology Transfer at the Institute of Telecommunications and Multimedia Applications (ITEAM), the most prestigious research institute at UPV. Throughout his career, he has led as Principal Investigator more than 40 R&D projects. He has authored or co-authored more than 60 scientific papers in prestigious journals of signal processing and acoustics, and more than 160 papers in conference proceedings. His research interests encompass audio, acoustics, and multimedia, with current focus on digital audio processing in spatial sound, sound perception, signal processing algorithms, real-time multimedia systems development, and machine learning and artificial intelligence applications to sound and image. Dr. Lopez currently serves as vice-chair of the Spanish Section of the Audio Engineering Society.

Table 25

Jose M. Mossi received his Ingeniero de Telecomunicación degree from Universidad Politécnica de Madrid in 1991, and his Ph.D. from Universitat Politècnica de València (UPV) in 1998. Since 1991, he has been a member of the Department of Communication at UPV where, from 1999, he is Profesor Titular He has participated in numerous national and European research projects and has published over 40 papers in international technical journals and at renowned conferences in the fields of transportation, medical, and sports image processing.

Table 26

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Descriptor: LYNRED Mobility Dataset Multimodal Detection Subset (LYNRED-MDS)
Jul 2, 2026Loïc Arbez, Jessy Matias, Xavier Brenière +2Autonomous Driving Datasets Autonomous Driving Benchmarks
Thoth · INRIA center at University Grenoble Alpes, 38330 Montbonnot-Saint-Martin, France · Lynred, 38113, Veurey-Voroize, France +2
Multimodal Emergency Vehicle Classification via Audio-Visual Transformers and Knowledge Distillation
Sep 15, 2026Vijay John, Amar DabajaMultimodal Robustness Multimodal Classification
Department of Mathematics and Computer Science Lawrence Technological University, Southfield, MI, USA
Attention-Augmented YOLOv8 with Ghost Convolution for Real-Time Vehicle Detection in Intelligent Transportation Systems
Apr 22, 2026Syed Sajid Ullah, Muhammad Zunair Zamir, Ahsan Ishfaq +1Intelligent Transportation Systems Object Detection
School of Energy and Electrical Engineering, Chang’an University, Xi’an, China · School of Information Engineering, Chang’an University, Xi’an, China

Jul 2, 2026J/K move · Enter open · S save

Descriptor: LYNRED Mobility Dataset Multimodal Detection Subset (LYNRED-MDS)

Loïc Arbez, Jessy Matias, Xavier Brenière, Jocelyn Chanussot+1

Thoth · INRIA center at University Grenoble Alpes, 38330 Montbonnot-Saint-Martin, France · Lynred, 38113, Veurey-Voroize, France+2

Autonomous Driving Datasets Autonomous Driving Benchmarks Multispectral Object Detection Infrared Object Detection Object Detection

Open paper AskPDF arXiv

Current road safety systems primarily focus on minimizing post-collision damage. However, advances in algorithmic perception are shifting focus toward early collision prediction, especially in lowvisibility conditions like nighttime or fog, where thermal infrared sensing outperforms both human vision and RGB imaging. While available RGB-infrared datasets such as FLIR ADAS and LLVIP are good benchmarks, they mostly consist of clear weather and overly simple scenarios. In this article, we introduce the LYNRED-MDS: Multimodal Detection Subset, a subset of the LYNRED Mobility Dataset, comprised of 4000 RGB-infrared image pairs captured under diverse weather, lighting, and road conditions around Grenoble, France. Our dataset spans varied driving contexts (urban, rural, mountainous, etc.) and a vehicle fleet compliant with Western European standards. Thermal cross-dataset evaluation using a YOLOv8n baseline suggests that our dataset offers strong generalization potential for pedestrian detection in driving scenarios. By covering critical edge cases, our dataset supports the development of more reliable and deployable vision systems for advanced driver-assistance systems.

Abstract

Object Detection

Explore similar work

Descriptor: LYNRED Mobility Dataset Multimodal Detection Subset (LYNRED-MDS)

Multimodal Emergency Vehicle Classification via Audio-Visual Transformers and Knowledge Distillation

Attention-Augmented YOLOv8 with Ghost Convolution for Real-Time Vehicle Detection in Intelligent Transportation Systems