Geçtiğimiz son on yılda, Derin Öğrenme, bilhassa Evrişimsel Sinir Ağları (Convolutional Neural Network), Makine Öğrenmesi ve Derin Sinir Ağları’nın en hızlı büyüyen alanıdır. Birçok Derin Sinir Ağı arasında, günümüzde Evrişimsel Sinir Ağları görüntü analizi ve sınıflandırma amaçları için kullanılan ana araçların başında gelmektedir. Evrişimsel Sinir Ağı tabanlı modeller, yüz tanıma görevlerinde yüksek başarı performansı sergilemektedirler. İlgili modellerin başarı performansının yüksek olması oluşturulan mimariye ve tercih edilen hiper-parametrelere bağlıdır. Ek olarak, modellerin eğitildikleri veri setinin boyutu performans üzerinde büyük etkiye sahiptir. Bu çalışmada ana amacımız, afin dönüşümü (affine transform) yöntemi ile veri artırma işlemini gerçekleştirmek ve bu veri artırma tekniğinin Evrişimsel Sinir Ağlarına dayalı yüz tanıma sistemine etkisini analiz etmektir. Evrişimsel Sinir Ağlarına dayalı yüz tanıma sistemi Destek Vektör Makineleri ve K-En Yakın Komşu sınıflandırma algoritmaları ile uygulanmış, devamına iki algoritmanın performansı karşılaştırılmıştır. Çalışmamızda bütün deneyler Labeled Faces in the Wild (LFW) veri seti üzerinde gerçekleşmiştir. Elde edilen sonuçlar, uygulanan veri artırma tekniğinin, yüz doğrulama işlemi için %1.8 oranında, yüz sınıflandırma işlemi için %2.2 (Destek Vektör Makineleri) ve %2.5 (K-En Yakın Komşu) oranında artış sergilediği gözlemlenmiştir. Gerçekleşen bütün deneyler sonunda, yüz tanıma sistemi doğrulama işleminde %94.4 oranında doğruluk elde etmiştir. Sınıflandırma işlemlerinde ise sistem, Destek Vektör Makineleri algoritması uygulanarak %97.1, K-En Yakın Komşu algoritması uygulanarak ise %96.3 oranında başarı performansı elde etmiştir.

Detecting and identifying signboards on their route is crucial for all autonomous and semi-autonomous vehicles, such as delivery robots, UAVs, UGVs, etc. If autonomous systems interact more with their environments, they have the ability to improve their operational aspects. Extracting and comprehending textual information embedded in urban areas has recently grown in importance and popularity, especially for autonomous vehicles. Text detection and recognition in urban areas (e.g., store names and street nameplates, signs) is challenging due to the natural environment factors such as lighting, obstructions, weather conditions, and shooting angles, as well as large variability in scene characteristics in terms of text size, color, and background type. In this study, we proposed three stages text detection and recognition approach for outdoor applications of autonomous and semi-autonomous mobile robots. The first step of the proposed approach is to detect the text in urban areas using the "Efficient And Accurate Scene Text Detector (EAST)" algorithm. Easy, Tesseract, and Keras Optical Character Recognition (OCR) algorithms were applied to the detected text to perform a comparative analysis of character recognition methods. As the last step, we used the Sequence Matcher to the recognized text values to improve the method's impact on OCR algorithms in urban areas. Experiments were held on the university campus by an 8-wheeled mobile robot, and a video stream process was carried out through the camera mounted on the top of the mobile robot. The results demonstrate that the Efficient And Accurate Scene Text Detector (EAST) text detection algorithm combined with Keras OCR outperforms other algorithms and reaches an accuracy of 91.6%

Robust autonomous navigation systems rely on mapping, locomotion, path planning, and localization factors. Localization, one of the most essential factors of navigation, is a crucial requirement for a mobile robot because it needs the capability to localize itself in the environment. Global Positioning Systems (GPSs) are commonly used for outdoor mobile robot localization tasks. However, various environmental circumstances, such as high-rise buildings and trees, affect GPS signal quality, which leads to reduced precision or complete signal blockage. This study proposes a visual-based localization system for outdoor mobile robots in crowded urban environments. The proposed system comprises three steps. The first step is to detect the text in urban areas using the “Efficient and Accurate Scene Text Detector (EAST)” algorithm. Then, EasyOCR was applied to the detected text for the recognition phase to extract text from images that were obtained from EAST. The results from text detection and recognition algorithms were enhanced by applying post-processing and word similarity algorithms. In the second step, once the text detection and recognition process is completed, the recognized word (label/tag) is sent to the Places API in order to return the recognized word’s coordinates that are passed within the specified radius. Parallely, points of interest (POI) data are collected for a defined area by a certain radius while the robot has an accurate internet connection. The proposed system was tested in three distinct urban areas by creating five scenarios under different lighting conditions, such as morning and evening, using the outdoor delivery robot utilized in this study. In the case studies, it has been shown that the proposed system provides a low error of around 4 m for localization tasks. Compared to existing works, the proposed system consistently outperforms all other approaches using just one sensor. The results indicate the efficacy of the proposed system for localization tasks in environments where GPS signals are limited or completely blocked.