Enhanced YOLO and RAFT for Real-Time Detection and Depth Estimation of Multiple Fruits in Orchard Environments

In the modern fruit and vegetable industry, manual harvesting incurs significant costs.Harvesting robots can reduce labor costs and improve efficiency and economic benefits.However, challenges such as occlusion, uneven lighting, and adverse weather conditions complicate the rapid and accurate identification and localization of fruits and vegetables.To address these issues, we propose a solution that integrates an enhanced YOLO (You Only Look Once) network for real-time and accurate object detection with RAFT (Recurrent All-Pairs Field Transforms) for depth estimation.This combination significantly improves both the speed and accuracy of the detection and localization process.We conducted experiments on our own orchard dataset and public datasets, demonstrating good performance in terms of average precision, frame rate, and depth estimation error.The results show that our method is suitable for real-time orchard harvesting tasks and has practical value.Additionally, to enhance the model's generalizability, we extended our experiments to include a variety of other fruits and vegetables across multiple natural orchard environments, confirming robust performance under different conditions.