Automatic detection of kiwi fruit with support vector machines using LiDAR point cloud

With the advancement of automation in agriculture, determining the spatial locations of fruits has become crucial for precision agriculture in order to enable agricultural applications such as monitoring fruit health, yield estimation, and automated harvesting. In this thesis study, LiDAR point cloud data of kiwi trees were used to segment kiwi and non-kiwi points on the tree using the Support Vector Machine (SVM) algorithm. The study aimed to contribute to yield estimation in the agricultural sector by utilizing machine learning methods. The red, green, and blue color values obtained from the LiDAR data, along with the surface normal vector values calculated in the X, Y, and Z directions for each point, were used as input data for the SVM, and a training process was conducted for classification. In this study, the point cloud of a single kiwi tree scanned with a terrestrial LiDAR scanner was used. The point cloud data was divided into six parts, with two used for SVM training and four for testing. During the SVM training process, the Grid Search Optimization algorithm was utilized to determine the most suitable hyperparameters for the given data, and the training was performed accordingly. The accuracy values for the training process were calculated as 0.95 for Overall Accuracy, 0.88 for Precision, 0.86 for Recall, 0.87 for F1-Score, and 0.79 for Intersection over Union (IoU). The testing process was conducted using four parts of the kiwi tree's point cloud, and the average accuracy values for the same metrics were calculated as 0.95, 0.80, 0.86, 0.82, and 0.73, respectively.