Publication: Stereo-Kamera İle Üretilen 3b Model Doğruluğunun Analizi
Abstract
Bu çalışmada, mobil insansız kara aracına entegre edilen stereo kamera sensörü kullanılarak iç mekân ortamı taranmış ve elde edilen görsel verilerden üç boyutlu (3B) bir model oluşturulmuştur. Çalışmanın temel amacı, stereo kamera tabanlı modellemenin geometrik doğruluğunu değerlendirmek ve bu doğruluğu etkileyen başlıca etkenleri ortaya koymaktır. Model doğruluğunu değerlendirmek amacıyla iki aşamalı bir analiz yaklaşımı benimsenmiştir. İlk aşamada, çalışma alanında yer alan toplam 34 farklı nesneye ait boyutsal veriler (en ve boy) hem oluşturulan 3B model üzerinden hem de yüksek hassasiyetli Total Station cihazı ile ölçülmüş ve sonuçlar karşılaştırılmıştır. Bu karşılaştırmalar sonucunda, ortalama hata en yönünde ±0.0057 m, boy yönünde ise ±0.0097 m olarak hesaplanmış; bazı nesnelerde daha yüksek sapmaların görüldüğü tespit edilmiştir. Bu sapmaların, yüzey dokusu eksikliği, ışık yansımaları ve kamera görüş açısı gibi faktörlerden kaynaklandığı belirlenmiştir. İkinci aşamada ise modelin yüzey doğruluğu değerlendirilmiştir. Bu kapsamda toplam 24 yüzey (sol duvar, sağ duvar ve zemin) üzerinde yapılan analizlerde, ortalama RMSE ±1.43 cm, ortalama mesafe farkı ±15.68 cm ve standart sapma ±2.03 cm olarak bulunmuştur. Ancak bazı yüzeylerde ortalamanın üzerinde hatalar tespit edilmiş; özellikle dokusuz, geniş ve ışığa maruz kalan alanlarda stereo eşleşmenin zorlaşmasından kaynaklı derinlik sapmaları gözlenmiştir. Elde edilen bulgular, stereo kamera verileriyle üretilen 3B modelin santimetre düzeyinde hassasiyete modelleme yapabildiğini göstermekte; ancak model kalitesinin, yalnızca kullanılan sensörün teknik özelliklerine değil, aynı zamanda yüzey geometrisi, dokusal yapı, ışık koşulları ve araç hareket dinamikleri gibi dışsal faktörlere de bağlı olduğunu ortaya koymaktadır. Bu nedenle, stereo kamera tabanlı modelleme uygulamalarında, çevresel koşulların ve yüzey özelliklerinin dikkatle değerlendirilmesi gerektiği sonucuna varılmıştır.
In this study, a specific indoor environment was scanned using the stereo camera sensor integrated into a mobile unmanned ground vehicle, and a three-dimensional (3D) model was generated from the captured visual data. The main objective of the study is to assess the geometric accuracy of stereo camera-based modeling and to identify the key factors that influence this accuracy. A two-stage analysis approach was adopted for the evaluation. In the first stage, dimensional measurements (width and height) of 34 different objects within the study area were obtained both from the 3D model generated by the stereo camera and physically using a high-precision terrestrial surveying instrument. By accepting the data from the terrestrial measurements as reference, average errors of ±0.0057 m for width and ±0.0097 m for height were calculated, with larger deviations observed for certain objects. These deviations were attributed to factors such as textureless surfaces, light reflections, and limited camera viewing angles. In the second stage, a general surface accuracy analysis was conducted. A total of 24 surfaces (including left walls, right walls, and floors) were evaluated by comparing stereo camera data with reference measurements. The results showed an average RMSE of ±1.43 cm, an average distance difference of ±15.68 cm, and a standard deviation of ±2.03 cm. However, significantly higher errors were observed on certain surfaces, particularly on large, textureless, and well-lit areas, where stereo matching becomes more challenging. The findings demonstrate that the 3D model generated by stereo camera data is capable of producing centimeter-level accuracy in 3D modeling. However, model quality depends not only on the technical specifications of the sensor, but also on external factors such as surface geometry, texture, lighting conditions, and the stability of sensor motion. Therefore, in stereo camera-based modeling applications, these environmental and structural parameters should be carefully managed.
In this study, a specific indoor environment was scanned using the stereo camera sensor integrated into a mobile unmanned ground vehicle, and a three-dimensional (3D) model was generated from the captured visual data. The main objective of the study is to assess the geometric accuracy of stereo camera-based modeling and to identify the key factors that influence this accuracy. A two-stage analysis approach was adopted for the evaluation. In the first stage, dimensional measurements (width and height) of 34 different objects within the study area were obtained both from the 3D model generated by the stereo camera and physically using a high-precision terrestrial surveying instrument. By accepting the data from the terrestrial measurements as reference, average errors of ±0.0057 m for width and ±0.0097 m for height were calculated, with larger deviations observed for certain objects. These deviations were attributed to factors such as textureless surfaces, light reflections, and limited camera viewing angles. In the second stage, a general surface accuracy analysis was conducted. A total of 24 surfaces (including left walls, right walls, and floors) were evaluated by comparing stereo camera data with reference measurements. The results showed an average RMSE of ±1.43 cm, an average distance difference of ±15.68 cm, and a standard deviation of ±2.03 cm. However, significantly higher errors were observed on certain surfaces, particularly on large, textureless, and well-lit areas, where stereo matching becomes more challenging. The findings demonstrate that the 3D model generated by stereo camera data is capable of producing centimeter-level accuracy in 3D modeling. However, model quality depends not only on the technical specifications of the sensor, but also on external factors such as surface geometry, texture, lighting conditions, and the stability of sensor motion. Therefore, in stereo camera-based modeling applications, these environmental and structural parameters should be carefully managed.
Description
Citation
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
92
