Fruit Detection from Apple Orchard Using Point Cloud Data Nokta Bulutu Verisi Kullanılarak Elma Bahçesinden Meyve Tespiti

Creative Commons License

Günen M. A.

El-Cezeri Journal of Science and Engineering, vol.9, no.1, pp.253-265, 2022 (Scopus) identifier

  • Publication Type: Article / Article
  • Volume: 9 Issue: 1
  • Publication Date: 2022
  • Doi Number: 10.31202/ecjse.962269
  • Journal Name: El-Cezeri Journal of Science and Engineering
  • Journal Indexes: Scopus
  • Page Numbers: pp.253-265
  • Keywords: Apple detection, Classification, Feature extraction, Feature selection, Point cloud
  • Kayseri University Affiliated: No


© 2022, TUBITAK. All rights reserved.In precision agriculture applications, it is critical to determine the spatial location of fruits in order to perform agricultural management applications such as monitoring fruit health throughout the development process, automatic harvesting, and yield estimation. Due to the complex geometry of orchards, developing modern solutions in orchards is difficult. The majority of fruit detection research is based on image analysis. A photogrammetrically produced point cloud from a Fuji apple orchard was used in this study. To detect the spatial positions of apples, a new framework has been proposed. In the proposed framework, an Omnivariance-based approach was used to determine the most suitable neighborhood. After determining the most suitable size of neighborhoods, 30 2D and 3D geometric features were extracted from each individual point. Then, the features that best represent the data set were selected using the minimum redundancy maximum relevance method. In order to examine the effects of different features on apple detection, the related features were divided into six different groups according to their weight level and statistical and visual comparisons were made. According to the results of the classification process using a support vector machine, the use of 25 features (95.81% accuracy and 93.20% precision) provided the highest classification performance. It has been determined that the use of all or a limited number of features reduces the classification performance. In addition, 2D features were found to be as effective as 3D features.