Machine Learning Feature Extraction Based on Binary Pixel Quantification Using Low-Resolution Images for Application of Unmanned Ground Vehicles in Apple Orchards

Abstract

Deep learning and machine learning (ML) technologies have been implemented in various applications, and various agriculture technologies are being developed based on image-based object recognition technology. We propose an orchard environment free space recognition technology suitable for developing small-scale agricultural unmanned ground vehicle (UGV) autonomous mobile equipment using a low-cost lightweight processor. We designed an algorithm to minimize the amount of input data to be processed by the ML algorithm through low-resolution grayscale images and image binarization. In addition, we propose an ML feature extraction method based on binary pixel quantification that can be applied to an ML classifier to detect free space for autonomous movement of UGVs from binary images. Here, the ML feature is extracted by detecting the local-lowest points in segments of a binarized image and by defining 33 variables, including local-lowest points, to detect the bottom of a tree trunk. We trained six ML models to select a suitable ML model for trunk bottom detection among various ML models, and we analyzed and compared the performance of the trained models. The ensemble model demonstrated the best performance, and a test was performed using this ML model to detect apple tree trunks from 100 new images. Experimental results indicate that it is possible to recognize free space in an apple orchard environment by learning using approximately 100 low-resolution grayscale images.

Keywords

• machine learning (ML)
• unmanned ground vehicle (UGV)
• orchard
• binary image
• feature extraction
• ensemble model