Goran Paulin

Goran Paulin

Auto-generated profile, based on publicly available information

Goran Paulin

Faculty of Informatics and Digital Technologies, University of Rijeka, Rijeka, Croatia

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 January 2023 in Artificial Intelligence Review
Verified
Journal Article
Closed access
PDF file
GPGoran Paulin
Goran Paulin
Marina Ivasic-KosMarina Ivasic-Kos
Marina Ivasic-Kos


31 December 2025 in IEEE Open Journal of Intelligent Transportation Systems
Verified
Journal Article
Open access
Miran PobarMiran Pobar
Miran Pobar
GPG. Paulin
G. Paulin
Marina Ivasic-KosMarina Ivasic-Kos
Marina Ivasic-Kos
Aleksandar VorkapicAleksandar Vorkapic
Aleksandar Vorkapic
This work presents an automated, real-time approach for localized sea state estimation using a single camera mounted on a shipfs bridge. A labeled dataset of sea surface images, collected during regular operation of an overseas liner, is built and used to train deep neural networks to estimate sea state on the Beaufort scale from 1 to 8. To improve robustness and better capture operational variability, a substantially enlarged test set is constructed relative to previous work, enabling a more comprehensive evaluation under diverse navigation and environmental conditions. Additionally, to mitigate the scarcity of rarely occurring sea states in real-world operation, a synthetic training dataset is generated that simulates a wide range of sea and weather conditions while preserving key physical relationships to increase variability in illumination and wave appearance without degrading realism. We evaluated state-of-the-art convolutional and transformer-based architectures, including Resnet-101d, DeiT III, Swin transformer, XciT and CoAtNet. The impact of different synthetic.to.real training data ratios in both RGB and grayscale domains is systematically examined, yielding a 6% improvement in test accuracy and mean F1 score and a reduction of the maximum error from 7 to 3 Beaufort. Finally, a temporal voting framework that aggregates predictions over several consecutive frames further reduces the maximum error to 2 Beaufort and achieves 96% intra.1.class accuracy and an F1 score of 62%, substantially outperforming a baseline trained only on real data without temporal voting. Impact Statement – The maritime industry is increasingly adopting artificial intelligence to enhance operational efficiency, improve navigational safety and achieve sustainability goals. Ensuring navigational safety, as well as the environmental and economic efficiency of maritime operations, requires an accurate assessment of sea state. Traditionally, sea state is visually assessed using the Beaufort scale, which relates wind speed to sea surface state and classifies it into 13 classes (0-12). Since this method relies on visual observation, it is subjective and prone to human error, so automating sea state assessment using computer vision methods can provide an effective monitoring alternative. The models proposed in this paper achieved intra-1-class accuracy of 96% on a varied test set, demonstrating the effectiveness of this approach for robust sea state estimation.

8 September 2021
Verified
Conference Paper
Closed access
GPGoran Paulin
Goran Paulin
SSSasa Sambolek
Sasa Sambolek
Marina Ivasic-KosMarina Ivasic-Kos
Marina Ivasic-Kos
By using drones in search and rescue (SAR) missions, missing person detection is possible during or after the flight by analyzing the recorded material. However, person localization is equally important so that rescuers can approach the person in the shortest possible time. We propose a raycast method to determine both a person's location and the distance from the drone, using a sequence of monocular drone images. The proposed method has been tested in silico, using a custom-made procedural simulator, calibrated for windless and windy conditions. We concluded that multiple raycasting solves unreliable telemetry problems and that there is an optimal number of required iterations, depending on the telemetry noise of a specific drone.

15 April 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
GPGoran Paulin
Goran Paulin
MIMarina Ivasic-Kos
Marina Ivasic-Kos
Unmanned aerial vehicles are increasingly used in search and rescue missions, but person detectors trained on static real-world datasets generalize poorly across biomes, altitudes, and imaging configurations, and their performance degrades further when high-resolution imagery must be downscaled for inference. To address this, we propose the Adaptive Synthset System (AdaptSyS), a modular, data-centric pipeline that generates mission-specific synthetic datasets after mission start, fine-tunes an unmodified YOLO11s detector, executes SAHI-based tiled inference with an affirmative-voting ensemble, and performs distance-based coverage mapping onto a 3D terrain model using ray casting. AdaptSyS uses background crops from the first UAV flight and 3D crowd/ragdoll simulation to render a 3000-image adaptive synthset at 384 px resolution, and completes synthset generation, fine-tuning, and deployment within 46 minutes of mission initiation, satisfying typical SAR golden-hour constraints while remaining compatible with lowcost RGB-only platforms. A preliminary study on SARDv8, HERIDAL, and VisDrone2000 confirms strong domain dependence of YOLO detectors and shows that avoiding downscaling via cropping and SAHI tiling is critical, with 384 px crops yielding in-domain recall up to 0.932 and motivating ensembles across detectors and crop sizes. In an ablation study on HERIDAL, a model fine-tuned on an adaptive synthset generated with a Digital Content Creation method (AS-DCCv4) attains slightly lower standalone recall than a model fine-tuned on the real SARDv8-384 dataset (0.482 vs. 0.519), but produces 3.5× fewer false positives and contributes more unique true positives to the ensemble, demonstrating that a well-designed synthset can surpass a related real dataset in cross-domain benefit. When combined with the default YOLO11s in an affirmativevoting ensemble and evaluated with SAHI on full-resolution HERIDAL images, the adaptivesynthset ensemble achieves recall of 0.746 compared to 0.679 for the SARDv8-384 ensemble, while keeping the false-positive load operationally manageable, thus answering the research question affirmatively by showing that the adaptive synthset method can be deployed within 1 hour and can deliver equal or superior detection performance under new conditions. Finally, a distance-based mapper calibrated on the same adaptive synthset links detector range to effective ground coverage, quantifying the gap between nominally overflown area and terrain that is actually searchable at a suitable scale, which is essential for planning and evaluating UAV SAR missions.