Prof. Dr. Tankut Acarman danışmanlığındaki tezler
14 tez · Galatasaray University
Adlandırılmış varlık ile ad aktarması çözümleme
Internet and computers has become more and more important in human life. Especiallywith Web 2.0, the permission given to surfers of not only reading but also changingthe content has encouraged humans to change their habits of explorations. Nowadaysmankind use internet instead of encyclopaedias and books. This undeniable growth ofdigitalism put forth the importance of automatically processing of the data. Automati-cally processing the content means turning unstructured data into structured data. Byunstructured data we connote human language – natural language data. This treat-ment is necessary for computers to understand the natural language content. NaturalLanguage Processing (NLP) is the discipline behind this process. NLP is a subcategoryof artificial intelligence and computer linguistics.Named Entity Recognition (NER) and Word Sense Disambiguation (WSD) is two ofthe major tasks of NLP. NER is the classification and extraction process of word(s)considered significant in a text. This significant word(s) can differ according to field.For example, this entities may be percentages, dates as well as person names, locationnames and company names, etc. WSD is an open problem in NLP. It consists ofidentifying the sense of a word, when having multiple meaning, in a sentence. WSDtries to identify litteral expressions of a word, not figurative expressions. FigurativeLanguage Processing is the study similar and all but subfield of WSD. FigurativeLanguage Processing concentrates on determining figurative expressions except litteralexpressions.Our project is based on metonymy recognition and resolution through named entityrecognition. Metonymy is a figure of speech which consists by using a concept b to referto concept a, without intending analogy. The existing methods of metonymy resolutiondepends on supervised and unsupervised methods as well as statistical approaches. Thecommonly used approaches are catching the Selectional Restriction Violations (SRVs)and deviations from grammatical rules. We consider our project having three parts.First part is to pre-process the given text. Pre-processing is necessary for further treatment. Pre-processing consists of lemma-tization, part-of-speech tagging, NER tagging, dependency tagging and WSD treat-ment. The second part is metonymy recognition, in other words detections of possiblemetonymies. Metonymy recognition is realized via named entities' SRVs. It is doneby a rule based algorithm. The last and the third part is metonymy resolution whichconsists of determining metonymic relation.
Araçlar arası ağlar için kavşak yönlendirmeli taşıma protokolü
The smart city applications attract attention more and more in today's world, and vehicular network is one of the big and important part of them. VANET provides many safety and infotainment applications with DSRC (Dedicated Short-Range Communications). Since VANET's structure is ad-hoc based, it lacks the standard infrastructure of the classical computer networks. Some of the applications in VANET requires multi-hop routing between distant nodes in network. However, reliable routing is a major problem in VANET, because of random distribution and high mobility of the nodes. This thesis intends to provide a reliable routing algorithm in urban areas for VANET. With the help of GPS devices and digital maps, the proposed algorithm uses the network nodes at road intersections to achieve a successful geo-routing. In order to verify effectiveness of the proposed algorithm, two simulation studies are conducted. While SUMO is used to simulate urban city environment and vehicle behavior, NS-3 is used to simulate the vehicular network. MAC and Physical Layers of the network are implemented as IEEE 802.11p standards. The Network and Transport Layers is implemented as ETSI ITS standards. Within this context, the first results show that a reliable routing can be achieved with a simple digital map and even cheap GPS devices. The second study indicates that the proposed algorithm has affectively more success than some well-known routing algorithms.
Sürüş olayları veri setinde boyut indirgeme tekniklerinin karşılaştırılması
When investigating on problematical and indefinite areas with data exploring tools such as machine learning or DM algorithms, weight of data attributes effecting classification result is generally unknown issue. Using entire feature set might cause the low classification success. Dependency existance among features, (near) zero variance features, outlier and missing data on feature may harm classification accuracy. To increase classification success and learn feature effect on classification, dimension reduction techniques such as feature subset selection and feature extraction are used. Feature subset selection and feature extraction are the two different applied methods for reducing the dimension set. While feature subset selection methods is focusing to find the most important features that affect the classification result, feature extraction methods are dealing with the creating of new attributes as a linear or non linear combination of initial feature set. Both methods are used on the investigation of classification, clustering and regression problems. Feature extraction methods sacrifice the explanation of the problems, when they combine the existent features to create new ones. On the other hand, features subset selection methods help to pick the most important features by ordering attributes according to their ranking methods. If classification researches are not satisfying or contribution of the attributes that affecting the classification is not known deeply, both methods can be used to understanding the importance of the attributes, and increase classification accuracy.
Kamera ve araç-araç haberleşmesi birleşimli gerçek zamanlı araç takibi
Nowadays autonomous cars have already started to be used in our daily lives. They are the new generation vehicles, which are able to operate themselves under certain conditions without the need of a driver's response. They can follow the lane markers and keep themselves inside the lane, track the vehicles around and decide and perform a break or even change the present lane where necessary. Driver assistance systems are the underlying technology of these vehicles. It enables them to analyze the road and traffic conditions in order to take these kinds of necessary actions. This makes the task of detection of objects at the road such as vehicles, lane markers, guardrails and even pedestrians walking by highly important. For this reason, it is also very substantial to be able to operate in real time conditions. Current studies focused on this area are mostly based on vision sensor, LIDAR, microwave sensors etc. However, they alone suffer from conditions such as high variety of targets, lighting fluctuations, short range, requirement of high process power, cost etc. In this work, we introduce a monocular camera based system, which is enhanced with the IEEE 802.11p vehicle-to-vehicle communication standard. By the help of IEEE 802.11p we aim to compensate the cons of the vision sensor systems which are mentioned above. Sensor measurements are fused using the particle filter algorithm. Due to importance of real time requirement, we also made tests on the GPU accelerated NVIDIA Tegra Jetson TK1 development board and compare with the CPU results for future development.
Geçici pencerede çağrışımlara dayalı çoklu nesne takibi
In this thesis, multiple-object vehicle tracking system that generates tracklets by solving the min-cost max flow problem of affinity between detected objects of consecutive frames are proposed. Recent performance enhancement in runtime and detection accuracy of Convolutional Neural Networks and their capability is object detection created the tracking-by-detection paradigm. Tracking-by-detection is the approach of data i.e. identity, association between individual frames of a sequence. The proposed tracking system is targeted to autonomous driving applications, with being able to track on non-stationary scene recordings, i.e. from a moving ego vehicle. Object detection task is performed using an Neural Network Ensemble consisting of 3 Faster R-CNN Inception ResNet v2 networks, that are trained on ImageNet dataset and fine-tuned for KITTI Object Detection Dataset. A method is proposed for combining the bounding boxes generated from each of the Convolutional Neural Network. Efficient usage of the processing resources for autonomous driving that should satisfying the requirements of computationally complex tasks such as localization, object detection, occupancy grid update, sensor-fusion, trajectory planning etc. are respected by generating tracklets on a temporal window of three frames. Also data association is done by solving min-cost flow of sparse affinity network rather than considering all possible assignments, that solves memory and computational constraints. Lighter version of the method is also presented where the temporal window is reduced to two consecutive frames and data association is done by solving linear sum assignment problem, minimum weight matching of a bipartite graph introduced by the Hungarian Algorithm. For evaluation of the trackers KITTI Object Tracking Evaluation 2012 dataset and 'Car' class is used. KITTI Object Tracking dataset consists of 21 training sequences with 8.008 frames and 29 testing sequences with 11.095 frames. Frames were recorded at 10 FPS from a camera mounted on the ego vehicle. All sequences have varying number of objects and lengths with their unique motion scenarios. In our evaluation study, the following metrics are adopted: Recall, Precision, F1-Measure, False Alarm Rate, False Positives, False Negatives relating to the object detection task and Runtime, widely used CLEAR MOT metrics like Multiple Object Tracking Accuracy and also Fragmentation ,ID-switch, Mostly-Tracked and Mostly-Lost for MOT evaluation. with our model performing second on MOTA, MT, ML metrics compared to the state-of-the-art online MOT methods. It showed less than half of the reported IDS from the best MOTA metric and lower FRAG, while working 6 times faster, with mean runtime at 20 Hz.
Varlık ismi tanımlama üzerine kosinüs benzerliği özelliğinin farklı örneklemlerde değerlendirilmesi
Today, social media is a huge part of our world and it continues to grow exponentially. Enormous content is being created with these platforms and it draws the attention of people for personal and professional levels. However, extracting meaningful information from this volume of content with human capabilities is not possible. Machine learning approaches are used to solve this problem with the help of computer power. Natural Language Processing (NLP) is a branch of artificial intelligence which is focused on interacting humans and computers using the natural language. By the aid of machine learning, NLP can achieve tasks from text such as tokenization, classification, sentiment analysis, Named Entity Recognition (NER). These tasks are producing successful results for well-structured texts like newspapers, articles, and books yet, working with unstructured texts from social media are challenging. These types of texts contain emoticons, abbreviations, grammar mistakes, and code-switching making data unpredictable and dirty. Twitter is one of the most popular microblog among social media platforms. It provides texts which are publicly posted and contains topic-specific opinions. It is a valuable source for collecting data. On the other hand, the content is unstructured because of character limitation and casual writing. In this study, we present an NER system and we evaluate baseline classifiers for unstructured texts. We develop cosine similarity feature and we evaluate and test each classifier subject to different combinations of features with cosine similarity. Our experimental results show that the presented system is reached at 74% level in precision, 68% in recall and 67% in F1 (micro average), respectively for Named Entity rEcognition and Linking (NEEL) 2016 Challenge dataset. The corpus is created from Twitter. In addition, we evaluate our system with 2 different datasets with different label distribution and types. One dataset is coming from a startup company named Oxtractor. It has 3 label types; "Person", "Organization", and "Location". Also, we present dataset which is labeled manually from specific topics of tweets. It has 7 types of the label; Person", "Thing", "Organization", "Location", "Product", "Event", and "Character". We compare the prediction results and try to find the best fitting classifiers and feature sets. Logistic regression, SVM, and Random forest are producing the highest results with cosine similarity feature. The results are obtained with different feature sets shows that supportive features for cosine similarly do not impact the results significantly. The diversity of named entity is distinctive when working with cosine similarity feature. Finally, we compare prediction results with different testing/training split ratios for the manually labeled dataset from 90/10 to 50/50. The cosine similarity feature does not affect the split ratio remarkably. Keywords: Named entity recognition, Information Extraction, Twitter, Word embedding, Classification, Machine learning, Cosine Similarity.
Akıllı araçlar için yörünge planlama
Scientific progress is an ongoing process and humans always have new ideas for new inventions. Although, scientific process is cumulative, you must invent the tire before inventing the car. Currently, mankind have past that point and looking towards to new challenges. One of the new ideas is cars that drive themselves, in a more formal and general way, the 'autonomous ground vehicles'. Autonomous vehicles in city traffic have been a dream for a long time. However, this great idea comes with its own set of problems. There are many parts of autonomous driving such as scene understanding, path planning. Many methods have been studied and developed to solve its problems. In this thesis, a novel path planning algorithm and results of extensive experimentation on it are presented. The algorithm runs real-time and is an alternative to well known existing algorithms. The experiments are done in a scalable custom environment consisting of a custom-built vehicle, desktop computer, a camera and an open source marker library.
Kara Taşıtlarının Diğer Yol Nesneleri ile Hücresel Haberleşme Teknolojisinde Kaynak Seçimi ve Benzetimi
Vehicular Ad-Hoc Networks is an emerging technology and research area since it promises fast and lossless communication among vehicular nodes. Vehicular to Everything communication targets data sharing between vulnerable road users and cars with mobile network assistance soon. Despite the nature of vehicular mobility and density, V2X requires transmitting massive data over a long range and being lossless. LTE is the most promising infrastructure for cellular V2X messaging, but it is still a developing technology requiring more research. To do so, a simulator environment is a fundamental need. In this thesis study, the literature review is conducted to elaborate on the LTE-V2X characteristics and performance evaluations in simulation environments. Moreover, a customized network simulator coupled with a road simulator is built to enable scalable simulation of various traffic scenarios. A set of physical channel configuration parameters is evaluated to maximize the performance of a fast and lossless ad-hoc communication in road traffic. Performance related parameters are listed and explained in the proposed system scheme. Resource selection techniques of C-V2X are reviewed and discussed. Our evaluation and test study illustrates that the resource selection algorithm plays a crucial role according to the calculated packet reception rates. Additionally, besides the prior public simulation studies, which conducted to seek the performance of C-V2X, are utilized to develop an NS-3 based open source C-V2X simulation environment with the geo-based resource selection algorithm implementation as a contribution along with comparative simulation studies. The proposed simulation environment introduces scalable and configurable system level and link level simulations in LTE vehicular networks.
Mobil lidar nokta bulutları için bir nesne ayırt etme yaklaşımı
LiDAR sensors are used to percept a scene by its 3D point cloud data. But LiDAR sensor measurement data are not homogeneous, they have uneven density and subject to occlusion problems. And the technical challenge lies along object segmentation because well-known clustering methods can not perform well with stand-alone 3D point cloud data. In this study, an approach is proposed by defining two classes of objects that are ground and nonground objects. The points are labeled ground and nonground points for DBSCAN based method. Then, sensitivity is adjusted for each class to improve the segmentation performance of ground and non-ground objects. For the CCL based method, three bird's eye view matrices are derived from point cloud data. The first matrix is used to separate ground and nonground areas and to merge other two matrix results. The ERCCL algorithm applied with a smaller sided cell for segmentation of ground objects and a bigger sided cell for segmentation of nonground objects. The presented approach is benchmarked with a DBSCAN based method and a CCL based method. The presented method outperforms the DBSCAN method by 4% and 13% and outperforms the CCL based method by 13% and %20 in v-measure and completeness clustering metrics, respectively. Overall, an approach is proposed to segment 3D terrains in an urban area with different segmentation methods. Keywords: 3D object segmentation, clustering, LiDAR, DBSCAN
Otonom sürüş için sensör birleştirme yaklaşımı
In this thesis, a sensor fusion approach between camera and lidar sensors is proposed to overcome shortcomings of single sensor setups. For camera sensor, a deep learning model, a multitask network is proposed to localize dynamic and static traffic objects and segment drivable area and lane lines. Dynamic objects include cars, motorcycles, bicycles, buses, trucks and pedestrians, static traffic objects include color classified traffic lights and traffic signs. This proposed multi-network achieves multiple tasks by itself faster and in real-time as normally there would be separate network models for each of these tasks working on parallel. The multi-network proposed is trained and tested with Berkeley Deep Drive 100K dataset. Evaluation results show that the proposed method is the fastest multi-network on the dataset with 47.62 FPS. Around multi-networks, proposed model has the second place on drivable area segmentation and lane line detection. Dynamic object localization performance of the network is state-of-the-art with 40% performance increase compared to other models. For Lidar sensor, 3D object detection and Lidar ground plane detection are achieved by already proposed methods. Finally, the detections from two sensors are fused by proposed fusion algorithms and results are evaluated with KITTI dataset. The proposed fusion approach exceeds the performance of Lidar-only methods up to 9.87% category wise. Comparison with two different fusion approaches also show our superior performance. Proposed sensor fusion methods can work in real-time by taking only 36 milliseconds, nearly 33 frames per second.
Türkçe tıbbi veri setlerinde ICD ve ATC sınıflandırmaları için NLP modellerinin karşılaştırmalı analizi
This study incorporates various methodologies to classify the Turkish medical texts with the aim of identifying ICD (International Classification of Diseases) & ATC (Anatomical Therapeutic Chemical) codes, leveraging various NLP approaches. The primary challenge involves comprehending the context and accurately categorizing medium to long documents into correct classes. Another challenge encompasses the acquisition of a labeled dataset with all categories, given the limited data resources in Turkish, which is a low resource language. Two distinct datasets are acquired for this study. The first dataset, which focuses on specific ICD-CM-10 C-Types, consists of medical summaries in Turkish language and was externally sourced. The second dataset, which is a new dataset including drug manuals covering all ATC types in Turkish is curated. After dataset creation, text processing has been implemented to obtain better performance results. SVC (Support Vector Classifier), FastText and various BERT models, such as BERTurk Uncased, BERTurk Large, ConvBERTurk, ElectraBERTurk and DistilBERTurk are used to classify the documents. Hyperparameter tuning is also applied to harness the potential of BERT and FastText, resulting in a notable 86% overall F1-Score on the Turkish medical summaries dataset and DistilBERTurk, the Turkish version of DistilBERT, resulted in 92.3% overall F1-Score on the same dataset. The most notable accomplishment, however, was for the Turkish Drug Manuals ATC code dataset, where the performance results, 96% F1-score using FastText, 95.9% F1-score using BERTurk Uncased and 94% using SVC models are obtained. These results highlight the effectiveness of the used approaches in Turkish text classification.
Bağlantılı ve otonom araçlarda siber güvenlik çalışması
The rapid development of autonomous vehicle technologies has made the security requirements of these systems more complex and has made it necessary to take security measures against cybersecurity threats in in-vehicle communication networks. In this thesis, a literature review is presented that comprehensively examines academic and industrial studies in the field of cybersecurity for autonomous and connected vehicles. The main purpose of the study is to address anomaly detection methods developed to detect security vulnerabilities in autonomous vehicles and to comprehensively present academic studies evaluating the role of these methods in determining security vulnerabilities. The differences between academic and industrial approaches in the field of cybersecurity are discussed in detail and an overview is provided regarding developments in the fields of generative artificial intelligence, machine learning, artificial intelligence and anomaly detection. In addition, various types of attacks such as fuzzy, malfunction and flooding, which are widely covered in the literature on cybersecurity vulnerabilities in autonomous vehicles and carry potential risks, are discussed and the prevention methods used against such attacks are evaluated. In the continuation of the study, various machine learning algorithms are divided into different test percentages and run on the relevant data set and presented in tables. These analyses were conducted using the In-Vehicle Intrusion Detection System Dataset, and a detailed assessment is provided on effective measures that can be taken against cyber-attacks. This study aims to make a comprehensive contribution to cyber security research in the field of autonomous and connected vehicles, and to constitute an important reference in the literature on the detection of security vulnerabilities in autonomous vehicles.
Sosyal medya verilerinin zaman-mekansal temellere göre makine öğrenimi
In recent years, machine learning algorithms have greatly revolutionized the methods used to analyze vast quantities of social media data and extract valuable insights from it. Researchers can leverage algorithms to enhance their understanding of human behaviors, reactions, and emotions, particularly in the context of natural disasters such as earthquakes. Analyzing social media data in terms of both space and time is extremely important. Utilizing machine learning on social media data related to earthquakes in a spatio-temporal manner allows for prompt interventions and allocation of resources, enabling early identification and rapid reaction by evaluating geo-tagged postings and real-time information sharing. It improves communication among individuals or groups involved and supports the ability to endure and recover from natural disasters by offering valuable and significant data on analyzing emotions, long-term strategies for recovery, and geographical patterns of damage. This study exploits crowdsourcing via social media data to extract information about emergency situations and needs after the earthquake. Using the semi-supervised method, the data has been labeled as either rescue or nonrescue to reach a high level of accuracy in detection. After the earthquake, rescue situations are detected on a spatial and temporal basis, along with location and time information provided by tweets. Two destructive earthquakes of magnitudes of Mw 7.7 and Mw 7.6 occurred on February 6, 2023, in the southeast of Turkiye. 53.537 people died, 107.213 people were injured, and several buildings were damaged. A total of 2.5 million tweets related to these earthquakes were collected from February 6 to February 28, 2023, through the X platform. For labeling purposes, nine BERT language models that are based on attention and transformers were used. Supervised learning methods, including logistic regression, support vector machines, decision trees, multinomial Naïve Bayes, and XGBoost, were applied to assess the precision of the labels and perform classification. Furthermore, the data set was processed with deep learning methods: convolutional neural networks, deep neural networks, and long short-term memory. A timely and proper response to delivering efficient solutions is possible only when the requests for assistance, rescue, and emergency are promptly and accurately understood. In this thesis, we determined the key terms of each data set through an extensive study of its spatio-temporal dynamics, allowing us to identify urgent supplies and use protection at the appropriate time quickly and clearly. The accuracy of data toward the detection of rescue and non-rescue situations is compared, and keywords on a spatio-temporal basis are extracted to determine hazard situations and emergency needs for coordination purposes. Deep learning and BERT models for detection of rescue and non-rescue classes reach a level of 0.8912 and 0.9792 in recall, respectively. This study highlights the vital importance of machine learning and deep learning in extracting valuable and applicable insights from social media data, especially in urgent scenarios like natural disasters. It achieves this by providing a thorough comprehension of the changing patterns that occur after the earthquake in Turkiye, incorporating both spatial and temporal factors into the analysis. The results demonstrate the effectiveness of the models in classifying microblogs connected to disasters.
Verisetinin artırılması ve interaktif modüllerin derin öğrenme mimarisine entegre edilmesi ile küçük nesne tespiti
Recently, image processing, video processing applications have made significant progress via Graphical Processing Unit emerging technology and has been effectively used by several industries such as robotics, augmented reality, autonomous cars, and surveillance systems. Object identification is a fundamental challenge in computer vision, as it serves as the basis for several sophisticated applications. By accurately detecting and localizing objects, computer vision systems may perform complex tasks such as object identification, semantic comprehension, and scene understanding. Driven by improvements in deep learning algorithms and the availability of big annotated datasets, object detection has made tremendous strides over time. Deep neural networks, in particular CNNs, have exceeded earlier techniques relying on manually created features and classifiers. CNNs have demonstrated impressive object detecting capabilities, offering greater accuracy.