Developing a real-time ai fruit detection model for robotic agriculture.

Abstract

The global agricultural sector faces growing pressure to meet increasing food demand while also dealing with reduced labor availability and the need for more sustainable and data driven farming practices. Real time fruit detection is a key requirement for robotic harvesting systems, yet many existing approaches struggle with changing environmental conditions, occlusions, and limited dataset diversity. This study develops and evaluates a real time fruit detection model based on the YOLOv4 deep learning architecture. The model was trained on a subset of the Google Open Images Dataset that contains eight fruit categories. A quantitative experimental design was used, which included image preprocessing, model training, optimization of learning parameters, and validation using standard evaluation metrics such as precision, recall, F1 score, mean average precision, and inference speed measured in frames per second. The model reached a mean average precision of 0.889, an average precision score of 0.89, an average recall of 0.85, and an inference speed of about 45 frames per second on a graphics processing unit. These results confirm that the model can operate in real time. Additional robustness tests under conditions that include variable lighting, partial occlusions, and cluttered backgrounds further showed that the model can perform reliably in simulated field environments. The study therefore provides a high accuracy and low latency fruit detection model that is suitable for integration into robotic harvesting systems. The work also supports the broader goals of precision agriculture and improved food security.