Wyniki wyszukiwania - BazTech

1

AIRIS: a real-time face and object detection system for threat monitoring

Urbaniak Ilona Anna, Kosek Wiktoria Maria, Kowalska Alicja Maria

Technical Transactions

|

2026

|

Vol. 123, iss. 1

art. no. e2026007

EN

This study presents AIRIS (Advanced Intelligent Recognition & Interception System), a real-time personal security monitoring platform integrating computer vision and artificial intelligence for mobile threat detection. The system is based on a three-layer architecture comprising adaptive face detection, temporal tracking, and hazardous object recognition using deep learning models. The main contribution lies in system-level integration and engineering validation under realistic deployment constraints. Individual identification combines embedding-based recognition with position-based tracking, while temporal persistence algorithms assess presence duration to identify potential risks. The implementation employs multithreaded processing and graceful degradation mechanisms to ensure reliable real-time operation in a wearable–mobile configuration. Experimental evaluation demonstrates 87% trial-level detection success for hazardous object presentation trials, 91% alert correctness, and processing throughput of 5–10 FPS with 120–180 ms latency.

2

Hoofed animal detection in UAV thermal images using Balanced Random Forest and CNN features

Włodarczyk Dorota, Jóźków Grzegorz

Reports on Geodesy and Geoinformatics

|

2025

|

Vol. 120

1--13

PL

Wildlife monitoring is vital to conservation efforts and the prevention of animal-related negative impacts on human activities and ecosystems. The use of Unmanned Aerial Vehicles (UAVs) enables data collection with no harm to wildlife and in difficult field conditions. This study proposes a method of detecting hoofed animals in UAV-acquired thermal images, addressing the challenges of low-resolution thermal imaging and the presence of other heated objects hindering simple temperature analysis and image segmentation. The proposed method uses machine learning algorithms and is designed to work with a limited size of training dataset. The method consists of an initial segmentation step that detects potential animals based on thermal and geometrical signatures, followed by classification using a Balanced Random Forest (BRF) algorithm. One of the key aspects of the proposed method is the use of geometric and thermal features along with multi-scale Convolutional Neural Network (CNN) extracted feature representations in BRF. The benefit of the BRF is its speed, little requirement regarding the amount of training data, and its capacity to work with an imbalanced number of objects in different classes. The dataset was collected during two UAV flights over a fenced enclosure with wild hoofed animals. The proposed approach showed high efficiency, achieving an overall accuracy of 90%. These results confirm the feasibility of UAV-based animal detection based solely on thermal images collected during the day and showing many other heated objects. The method provides a solution for wildlife monitoring, with potential adaptability to different species and further applications.

3

Dangerous items’ detection in surveillance camera images using Faster R ‑CNN

Omiotek Zbigniew

Przegląd Elektrotechniczny

|

2025

|

R. 101, nr 5

156--168

EN

The Faster R -CNN with different backbone networks was used to detect dangerous objects in the study. The best results were obtained for the ResNet152 backbone. The mAP value was 85%, while the AP level ranged from 80% to 91%, depending on the item detected. An average real -time detection speed was between 11 and 13 FPS. Both the accuracy and speed of the model allow it to be recommended for use in public security monitoring systems aimed at detecting potentially dangerous objects.

PL

W badaniach, do wykrywania niebezpiecznych obiektów wykorzystano sieć Faster R -CNN z różnymi sieciami szkieletowymi. Naj lepsze wyniki uzyskano dla sieci szkieletowej ResNet152. Wartość mAP wyniosła 85%, natomiast poziom AP wahał się od 80% do 91%, w za leżności od wykrywanego obiektu. Średnia prędkość wykrywania w czasie rzeczywistym wynosiła od 11 do 13 FPS. Zarówno dokładność, jak i szybkość modelu pozwalają rekomendować go do wykorzystania w systemach monitorowania bezpieczeństwa publicznego, mających na celu wykrywanie potencjalnie niebezpiecznych obiektów.

4

Analiza systemów rozpoznawania obiektów przy wykorzystaniu okularów do rozszerzonej rzeczywistości

Figura Jan, Kuźmiczuk Rafał, Badurowicz Marcin

Journal of Computer Sciences Institute

|

2025

|

Vol. 34

8--13

PL

Systemy rozpoznawania obiektów i urządzenia rozszerzonej rzeczywistości łączą świat wirtualny z rzeczywistym. Natu-ralnym krokiem wydaje się więc połączenie obu tych technologiiby tworzyć urządzeniawspierające różne aspekty życia. Niniejsza praca porównuje cztery modele wykrywania obiektów: Faster R-CNN ResNet-101 v1, YOLO v8s, SSD Mobi-leNet v2 i EfficientDet Lite 2 z perspektywy zastosowania w okularach rozszerzonej rzeczywistości.Wykorzystując sy-mulowane środowisko testowe, badane są zużycie zasobów, energooszczędność, precyzja i szybkość modeli podczas działania w czasie rzeczywistym. Uzyskane wyniki badań potwierdziły, że modele wykorzystującearchitekturę jedno-etapowąsą lepiej przystosowane do działania w czasie rzeczywistym, a najwydajniejszym okazał się YOLO v8s.

EN

Object recognition systems and augmented reality devices aim to bridge the gap between the virtual and real worlds. It is natural, then, to combine these two technologies to create devices that can assist us in various aspects of life. This paper compares four object detection models: Faster R-CNN ResNet-101 v1, YOLO v8s, SSD MobileNet v2, and EfficientDet Lite 2 in the context of their use with augmented reality glasses. Using a simulated test environment, we examined resource consumption, energy efficiency, precision, and speed of the models during real-time operation. The results confirm that models using single-stage architecture are more suitable for real-time operation directly on the device. Among the single-stage models, YOLO v8s proved to be the most efficient.

5

Applying AI-assisted image processing to detect and measure CO2 microbubbles for carbon storage

Le Dang Nguyen Hai, Le Nam Nguyen Hai

Inżynieria Mineralna

|

2025

|

Vol. 2, no. 1

755--761

EN

To address the challenges in the CO2 injection process, CO2 microbubble dispersion has been proposed as an alternative to traditional methods, such as miscible injection and water-alternating-gas (WAG) injection. This study presents an AI-assisted model for detecting CO2 microbubbles, powered by the YOLOv8 algorithm, renowned for its high-accuracy predictions. Conventional image processing techniques often struggle with detecting microbubbles, particularly in cases involving overlapping bubbles, variations in size, and low-contrast images, which can lead to inaccuracies in bubble identification and measurement. In contrast, YOLOv8’s advanced detection capabilities offer a more robust solution by precisely localizing and classifying microbubbles, even in challenging scenarios. The model’s performance was rigorously evaluated, demonstrating its effectiveness as a valuable tool for microbubble analysis. The detection images processed using YOLOv8 illustrate its ability to accurately detect and classify bubbles of varying sizes, generating precise bounding boxes around each identified bubble. This combination of data visualization and advanced detection techniques underscores the efficacy of YOLOv8 in microbubble analysis, enabling accurate measurement and detailed characterization of bubble size distributions—an essential factor in optimizing chemical engineering processes.

6

Fine-grained detection and segmentation of civilian aircraft in satellite imagery using YOLOv8

Panneerselvam Ramesh Kumar, Bandi Sarada, Doddapaneni Sree Datta Siva Charan

Informatyka, Automatyka, Pomiary w Gospodarce i Ochronie Środowiska

|

2025

|

T. 15, nr 2

5--12

EN

Detectionand segmentation of civilian aircraft from satellite imagery has significant importance in applications for air traffic management, surveillance, and defense. Yet, its visual confusions and lack of unification in recognition make it hard. This paper presents that by developing an efficient YOLOv8-based model for aircraft detection, classification, and segmentation within the FAIR1M-2.0 dataset. This proposed methodology involves dataset preprocessing and compatibility adjustments where the backbone used is CSPDarknet53 combining with the C2f module, which provides an efficient multi-scale representation, this happens to be the most critical requirement in distinguishing between among 11 unique categories of aircraft. Including the SAM model helps improve localization precision by achieving more accurate pixel-level segmentation. The present work effectively carried out an accurateclassification and described civilian aircraft, containing the enhanced detection and quantification capability appropriate for complex satellite-oriented aircraft analysis. These reasons make the work satisfy the fundamental requirement for very accurateidentification and evaluation of aerial images.The approach improves the accuracy and precision of aircraft classification over delicate satellite images, and thus is useful in operations for real-time surveillance and monitoring. Fine-grained classification and segmentation would then be able to effectively capture slight differences between aircraft types, which are now vital to the reliable management of airspaces. This work, therefore sets a good foundation for future development and advancement of high-resolution aerial analysis in diverse operational settings.

PL

Wykrywanie i segmentacja cywilnych samolotów na podstawie obrazów satelitarnych mają kluczowe znaczenie w zarządzaniu ruchem lotniczym, nadzorze oraz obronności. Ze względu na wizualne podobieństwa między różnymi typami samolotów oraz brak standaryzacji w rozpoznawaniu, jest to zadanie trudne. Niniejszy artykuł przedstawia efektywny model oparty na YOLOv8 do wykrywania, klasyfikacji i segmentacji samolotów w zbiorze danych FAIR1M-2.0. Zaproponowana metodologia obejmuje wstępne przetwarzanie danych i dostosowanie do zgodności, w którym wykorzystano CSPDarknet53 jako bazę, połączoną z modułem C2f, co zapewnia efektywną reprezentację wieloskalową–jest to kluczowy element przy rozróżnianiu 11 unikalnych kategorii samolotów. Włączenie modelu SAM poprawia precyzję lokalizacji, pozwalając na dokładniejszą segmentację na poziomie pikseli. Prezentowane badania pozwoliły na dokładną klasyfikację i opisanie cywilnych samolotów, zapewniając ulepszone możliwości wykrywania i analizowania obiektów na obrazach satelitarnych. Takie podejście znacznie zwiększa dokładność i precyzję klasyfikacji samolotów, co czyni je przydatnym w operacjach nadzoru i monitorowania w czasie rzeczywistym. Precyzyjna klasyfikacja i segmentacja umożliwia skuteczne rozróżnianie subtelnych różnic między typami samolotów, co jest istotne dla niezawodnego zarządzania przestrzenią powietrzną. Niniejsza praca stanowi solidną podstawę dlaprzyszłych badań nad analizą obrazów lotniczych w wysokiej rozdzielczości w różnych kontekstach operacyjnych.

7

Ore image target detection based on improved YOLOv5 network

Cai Gaipin, Luo Hui, Huang Jinruo, Yi Guotao

Gospodarka Surowcami Mineralnymi

|

2025

|

Vol. 41, iss. 2

177--196

EN

The existing target detection algorithms detect the ore on the conveyor belt after the crushing process with low precision and slow detection speed. This leads to challenges in achieving a balance between precision and speed, to enhance the detection precision and speed of ore, and in view of the problems of leakage, misdetection, and insufficient feature extraction of YOLOv5 in the task of ore image detection; this study presents a target detection approach relying on the CA attention mechanism (Coordinate attention for efficient mobile network design), the SIoU loss function and the target detection algorithm YOLOv5 combination of ore image particle target detection method. Integrating the CA attention mechanism into the YOLOv5 backbone feature network enhances the feature learning and extraction of ore images, thereby improving the precision of the detection model; the SIoU loss function is refined to boost the recognition precision of the network on ore images and address the shortcomings of the original loss function that fails to take angular loss, distance loss, and shape loss into account, thereby further improving the precision and speed of ore image detection. The experimental findings demonstrate that the AP value, value, and precision rate are improved compared with the pre-improved algorithm. The CA-YOLOv5 method is verified to be fast, effective, and advanced and provides a foundation for real-time target detection of ores on conveyor belts in subsequent intelligent mine production.

PL

Istniejące algorytmy wykrywania celu wykrywają rudę na taśmie przenośnika po procesie kruszenia z niską precyzją i niską szybkością wykrywania. Prowadzi to do wyzwań związanych z osiągnięciem równowagi między precyzją i szybkością, w celu zwiększenia precyzji i szybkości wykrywania rudy, a także ze względu na problemy z wyciekami, błędnym wykrywaniem i niewystarczającą ekstrakcją cech YOLOv5 w zadaniu wykrywania obrazu rudy; niniejsze badanie przedstawia podejście do wykrywania celu polegające na mechanizmie uwagi CA (Coordinate attention for efficient mobile network design), funkcji straty SIoU i kombinacji algorytmu wykrywania celu YOLOv5 w połączeniu z metodą wykrywania celu cząstek obrazu rudy. Zintegrowanie mechanizmu uwagi CA z siecią funkcji szkieletowych YOLOv5 usprawnia uczenie się funkcji i ekstrakcję obrazów rudy, tym samym zwiększając precyzję modelu wykrywania; funkcja straty SIoU została udoskonalona w celu zwiększenia precyzji rozpoznawania sieci na obrazach rudy i usunięcia niedociągnięć oryginalnej funkcji straty, która nie uwzględnia strat kątowych, strat odległości i strat kształtu, co jeszcze bardziej poprawia precyzję i szybkość wykrywania obrazów rudy. Wyniki eksperymentów pokazują, że wartość AP, wartość i wskaźnik precyzji są lepsze w porównaniu z wcześniej ulepszonym algorytmem. Metoda CA-YOLOv5 została zweryfikowana jako szybka, skuteczna i zaawansowana oraz stanowi podstawę do wykrywania celów rud na taśmach przenośnikowych w czasie rzeczywistym w późniejszej inteligentnej produkcji kopalnianej.

8

Czujniki pojemnościowe nowej generacji

Dozór Techniczny

|

2025

|

z. 4

28--28

PL

Czujniki pojemnościowe Turck serii BC/UC w stopniu ochrony IP67 łączą w sobie konwencjonalną pracę z cyfrową inteligencją i dodatkowymi danymi do monitorowania stanu.

9

Combining improved generative adversarial networks for end-to-end traffic object detection under complex illumination conditions

Liu Yang, Gong Zhe, He Yuyang, Li Weiqin

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2025

|

Vol. 73, nr 5

art. no. e155037

EN

The images captured by vehicle-mounted cameras in low-illumination environments have the problem of severe loss of detailed information. At the same time, the detection and recognition performance of traffic object detection algorithms is also influenced by factors such as object texture, movement speed, shooting angle, and occlusion. Under low-illumination conditions, the background of images is integrated with traffic objects, so the current object detection algorithms have relatively poor performance in detecting traffic objects under low illumination. In order to achieve low-illumination image enhancement without significantly reducing the reasoning speed of object detection algorithms and meanwhile improve the detection accuracy of object detection algorithms under low-illumination conditions, a multi-object detection model based on image enhancement, namely low-illumination enhancement and deep fusion-you only look once (LEDF-YOLO), is proposed. Firstly, based on the generative adversarial network (GAN) model, the direct-to-deep-generative adversarial network (DD-GAN) model is proposed to improve the effect of enhancing low-illumination images. Then, the fusion and parallel-cross stage partial bottleneck with two convolutions (FP-C2f) module and the transformer-spatial pyramid pooling fast (T-SPPF) module were designed to enhance and fuse multi-scale features. Finally, the network model of you only look once version 8n (YOLOv8n) was improved by introducing cross-hierarchical connections, making object localization more accurate. Experimental results on UA-DETRAC and self-made datasets showed that compared to the YOLOv8n algorithm, the LEDF-YOLO object detection method improved detection accuracy while maintaining the high real-time performance of the you only look once version 8n (YOLOv8n) algorithm.

10

Adaptive building engineering component extraction model based on DSOD

Lv Na, Yang Xuan

Archives of Civil Engineering

|

2025

|

Vol. 71, nr 2

345--361

EN

With the purpose to bring up the extraction efficiency and accuracy of building construction image component information, the dense block structure and loss function were proposed to optimize the deep supervised object detection algorithm, and an adaptive building construction component extraction model based on this algorithm was constructed. The improved depth-supervised target detection algorithm constructed by the study is validated and found to have an accuracy of 87.4% and a precision of 0.84, which is better than other comparative algorithms. The effectiveness of the adaptive extraction model of building components constructed by the research is verified, and it is found that the extraction error of the model is 9.8%, the value of the loss function is 0.2, and the satisfaction score of the experts is 8.8, and its extraction accuracy and efficiency are better than that of the other models, and it can satisfy the demand for the extraction of components of the construction project. In summary, it can be seen that the adaptive extraction model of building components constructed by the research has excellent information extraction performance, not only can it improve the efficiency of extracting engineering components, but it can also significantly enhance the decision support ability in construction management, optimize resource allocation, reduce risks, and improve the management efficiency of engineering projects. It has a positive contribution to the theory and practice of construction management discipline.

11

Metody zwiększania ilości danych termowizyjnych w uczeniu maszynowym

Sadzyński Piotr

Pomiary Automatyka Robotyka

|

2024

|

R. 28, nr 1

97--106

PL

Coraz częściej w przetwarzaniu i analizie obrazu termowizyjnego stosuje się uczenie maszynowe w kontekście rozpoznawania i identyfikacji obiektów. Niniejszy artykuł prezentuje wyniki badania wpływu augmentacji danych na efektywność uczenia maszynowego w kontekście analizy obrazów termowizyjnych. Wykorzystano publicznie dostępny zbiór danych FLIR ADAS, który zawiera etykietowane obrazy termowizyjne i obrazy z zakresu światła widzialnego. Badanie skupia się na wykorzystaniu konwolucyjnych sieci neuronowych, w szczególności architektury YOLOv8, do detekcji obiektów na obrazach termowizyjnych. Zbiór danych FLIR ADAS został poddany wstępnemu przetwarzaniu i augmentacji, a następnie wykorzystany do trenowania dwóch różnych modeli: jednego opartego na obrazach w skali szarości i drugiego - opartego na obrazach z zastosowaną paletą kolorów. Wyniki eksperymentu wskazują, że augmentacja danych może znacząco wpłynąć na efektywność modelu, a zastosowanie kolorów w obrazach termowizyjnych może w pewnych sytuacjach dodatkowo zwiększyć dokładność detekcji.

EN

Machine learning is increasingly being applied in the processing and analysis of thermal imaging for object recognition and identification. This article presents a study on the impact of data augmentation on the effectiveness of machine learning in the context of thermal image analysis. The publicly available FLIR ADAS dataset, which includes labeled thermal and visible light images, was used for this study. The research focuses on the use of Convolutional Neural Networks, specifically the YOLOv8 architecture, for object detection in thermal images. As part of the study, the FLIR ADAS dataset underwent preprocessing and augmentation, and was then used to train two different models: one based on grayscale images and another using a color palette. The results of the experiment indicate that data augmentation can significantly impact the effectiveness of the model, and the use of colors in thermal images may, in certain situations, further improve detection accuracy.

12

The concept of determining ground speed with limited access to GNSS signal on UAV

Kochan Jakub, Ważny Mariusz, Falkowski Krzysztof

Aviation and Security Issues

|

2024

|

No. 1

41--51

EN

This article presents an innovative approach to navigation using image recognition in situations characterized by limited access to GNSS (Global Navigation Satellite System) signals (signal interference). The presented system relies on image processing to define the characteristic edges of random objects. Subsequently, the actual ground speed of the moving object (UAV) is obtained based on changes in the object’s position. The article aims to show the potential of image recognition in navigational systems. The actual ground speed obtained by the image recognition can be used to correct the inertial navigation system.

PL

Niniejszy artykuł prezentuje nowatorskie podejście do nawigacji z wykorzystaniem rozpoznania obrazu w sytuacjach charakteryzujących się ograniczonym dostępem do sygnału GNSS (zakłócenia sygnału). Zaprezentowany system opiera się na przetworzeniu obrazu w celu zdefiniowania charakterystycznych krawędzi obiektu. Następnie na podstawie zmian położenia obiektu wyznaczana jest prędkość rzeczywista poruszającego się BSP. Ponadto w artykule przedstawiono potencjał zastosowania rozpoznania obrazu w systemach nawigacyjnych. Wyznaczona prędkość z użyciem rozpoznania obrazu może być wykorzystana do korekcji inercjalnych systemów nawigacji.

13

Object detection and multimodal learning for product recommendations

Selwon Karolina, Wnuk Paweł

TASK Quarterly : scientific bulletin of Academic Computer Centre in Gdansk

|

2023

|

Vol. 27, No 2

1--9

EN

This study showcases how deep learning can be applied to automated information extraction in fashion data to create a recommendation system. The proposed approach is an algorithm for recommending multiple products based on visual and textual features, ensuring compatibility with query items. The object detection model can detect many products across different garment categories. The study utilized public e-commerce datasets and trained models using deep learning methods. The compatibility model has shown promising results in automating recommendations of compatible products based on user interests. The study experimented with multiple pre-trained feature extraction models and successfully trained the object detection model for fashion article detection and localization tasks. Overall, the goal is to deploy the method to enhance its effectiveness and usefulness in providing a satisfying shopping experience for e-commerce users.

14

Estimation of absolute distance and height of people based on monocular view and deep neural networks for edge devices operating in the visible and thermal spectra

Gąsienica-Józkowy Jan, Cyganek Bogusław, Knapik Mateusz, Głogowski Szymon, Przebinda Łukasz

Annals of Computer Science and Information Systems

|

2023

|

Vol. 35

503--511

EN

Accurate estimation of absolute distance and height of objects in open area conditions is a significant challenge. In this paper, we address these problems and we propose a novel approach that combines classical computer vision algorithms with modern neural network-based solutions. Our method integrates object detection, monocular depth estimation, and homography- based mapping to achieve precise and efficient estimations of absolute height and distance. The solution is implemented on the edge device, which enables real-time data processing using both visual and thermography data sources. Experimental evaluation on a height estimation dataset prepared by us demonstrates an accuracy of 97.06\% and validates the effectiveness of our approach.

15

Vehicle detection and masking in UAV images using YOLO to improve photogrammetric products

Pargieła Karolina

Reports on Geodesy and Geoinformatics

|

2022

|

Vol. 114

15--23

EN

Photogrammetric products obtained by processing data acquired with Unmanned Aerial Vehicles (UAVs) are used in many fields. Various structures are analysed, including roads. Many roads located in cities are characterised by heavy traffic. This makes it impossible to avoid the presence of cars in aerial photographs. However, they are not an integral part of the landscape, so their presence in the generated photogrammetric products is unnecessary. The occurrence of cars in the images may also lead to errors such as irregularities in digital elevation models (DEMs) in roadway areas and the blurring effect on orthophotomaps. The research aimed to improve the quality of photogrammetric products obtained with the Structure from Motion algorithm. To fulfil this objective, the Yolo v3 algorithm was used to automatically detect cars in the images. Neural network learning was performed using data from a different flight to ensure that the obtained detector could also be used in independent projects. The photogrammetric process was then carried out in two scenarios: with and without masks. The obtained results show that the automatic masking of cars in images is fast and allows for a significant increase in the quality of photogrammetric products such as DEMs and orthophotomaps.

16

Intelligent agrobots for crop yield estimation using computer vision

Bini D., Pamela D., Mary Thusnavis Bella I., Shamia D., Prince Shajin

Computer Assisted Methods in Engineering and Science

|

2022

|

Vol. 29, no. 1-2 spec.

161--175

EN

The machine vision-based autonomous intelligent robots perform precise farm tasks suchas robot harvesting, weeding, pest or fertilizer spraying, monitoring, and pruning. Estimating crop yield is an essential assignment on a regional or federal scale. For a long timethe estimation measures were based on the statistics from manual counting of plants ina specific zone. The computer vision algorithms have addressed the technical drawbacksof the conventional image processing techniques and established an autonomous disciplineand yielded new approaches to crop planning. A method for quantitative assessment ofa tomato crop has been developed in this research using color thresholding in MATLAB using the RGB color model. Converting an RGB image to a grayscale image is one of thesteps involved in detecting red color in a taken image. After subtracting the two images,a median filter is employed to filter the noisy pixels to produce a two-dimensional blackand white image. The bounding boxes are used to label the binary digital images to detectrelated components, and the parameters of the labeled regions are computed to measurethe number of tomatoes in a crop. The obtained R2 correlation coefficient between thetomato berry counting algorithm and human counting was 0.98. Furthermore, the color ofeach pixel in the acquired image is evaluated by examining RGB values for pixel intensitiesin the obtained image. The performance of the berry counting algorithm was evaluated,and the technique was determined to have a high precision and recognition ratio of 96%.The research indicates that this technique may be used to estimate the crop yield, whichis helpful information for forecasting yields, planning harvest plans, and generating prescription maps for field-specific management strategies. The proposed model performedexceptionally well in estimating yield with each tomato (Solanum lycopersicum) crop.

17

Detection of unmanned aerial vehicles using computer vision methods: a comparative analysis

Hożyń Stanisław, Przybysz Marcin

Scientific Journal of Polish Naval Academy

|

2020

|

R. 61 nr 1-2 (220/221)

5--13

EN

Detection of small objects in the airspace is a crucial task in the military. In the era of today’s unmanned aerial vehicles (UAVs) technology, many military units are exposed to recognition and observation through flying objects. They are often equipped with optoelectronic warhead making a way to collect essential and secret data of the military unit. Modern technical solutions make it possible to implement some methods facilitating detection of flying objects. A lot of them utilize computer vision techniques based on image processing algorithm. Therefore, in this article, we present an analysis of the most promising algorithm for detection of small flying objects.

PL

W artykule przedstawiono analizę metod wykrywania bezzałogowych statków powietrznych wykorzystujących techniki widzenia komputerowego.

18

Object detection in the police surveillance scenario

Wilkowski Artur, Kasprzak Włodzimierz, Stefańczyk Maciej

Annals of Computer Science and Information Systems

|

2019

|

Vol. 18

363–-372

EN

Police and various security services use video analysis when investigating criminal activity. One typical scenario is the selection of object in image sequence and search for similar objects in other images. Algorithms supporting this scenario must reconcile several seemingly contradicting factors: training and detection speed, detection reliability and learning from sparse data. In the system that we propose a combined SVM/Cascade detector is used for both speed and detection reliability. In addition, object tracking and background-foreground separation algorithm together with sample synthesis is used to collect rich training data. Experiments show that the system is effective, useful and suitable for selected tasks of police surveillance.

19

License plate detection with machine learning without using number recognition

Ohzeki Kazuo, Geigis Max, Schneider Stefan

Annals of Computer Science and Information Systems

|

2019

|

Vol. 18

333–-340

EN

In autonomous driving, detecting vehicles together with their parts, such as a license plate is important. Many methods with using deep learning detect the license plate based on number recognition. However, there is an idea that the method using deep learning is difficult to use for autonomous driving because of the complexity in realizing deterministic verification. Therefore, development of a method that does not use deep learning (DL) has become important again. Although the authors have made the world's best performance in 2018 for Caltech data with using DL, this concept has now turned to another research without using DL. The CT5L method is the latest type, that includes techniques of the continuity of vertical and horizontal black-and-white pixel values inside the plate, unique Hough transform, only vertical and horizontal lines are detected, the top five in the order of the number of votes to ensure good performance. In this paper, a method to determine the threshold value for binarizing input by machine learning is proposed, and good results are obtained. The detection rate is improved by about 20 points in percent as compared to the fixed case. It achieves the best performance among the conventional fixed threshold method, Otsu's method, and the conventional method of JavaANPR.

20

Fast multispectral deep fusion networks

Osin V., Cichocki A., Burnaev E.

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2018

|

Vol. 66, nr 6

875--889

EN

Most current state-of-the-art computer vision algorithms use images captured by cameras, which operate in the visible spectral range as input data. Thus, image recognition systems that build on top of those algorithms can not provide acceptable recognition quality in poor lighting conditions, e.g. during nighttime. Another significant limitation of such systems is high demand for computational resources, which makes them impossible to use on low-powered embedded systems without GPU support. This work attempts to create an algorithm for pattern recognition that will consolidate data from visible and infrared spectral ranges and allow near real-time performance on embedded systems with infrared and visible sensors. First, we analyze existing methods of combining data from different spectral ranges for object detection task. Based on the analysis, an architecture of a deep convolutional neural network is proposed for the fusion of multi-spectral data. This architecture is based on the single shot multi-box detection algorithm. Comparison analysis of the proposed architecture with previously proposed solutions for the multi-spectral object detection task shows comparable or better detection accuracy with previous algorithms and significant improvement of the running time on embedded systems. This study was conducted in collaboration with Philips Lighting Research Lab and solutions based on the proposed architecture will be used in image recognition systems for the next generation of intelligent lighting systems. Thus, the main scientific outcomes of this work include an algorithm for multi-spectral pattern recognition based on convolutional neural networks, as well as a modification of detection algorithms for working on embedded systems.