Paddy field object detection for robotic combine based on real‐time semantic segmentation algorithm

• Published in: Journal of field robotics 2024-03, Vol.41 (2), p.273-287
• Main Authors: Zhu, Jiajun, Iida, Michihisa, Chen, Sikai, Cheng, Shijing, Suguri, Masahiko, Masuda, Ryohei
• Format: Article
• Language: English
• Subjects: Acceleration; Algorithms; Artificial neural networks; convolutional neural networks; deep learning; Frames per second; Harvesting; Machine learning; Microprocessors; object detection; Object recognition; paddy field; real‐time; robotic combine; Robotics; Semantic segmentation; Semantics; Sensors
• ISSN: 1556-4959; 1556-4967
• DOI: 10.1002/rob.22260

The development of robotic combine for rice harvesting has garnered worldwide attention in recent years. The robotic combine is capable of running along a designated path; however, it still requires human operator supervision due to the lack of object detection sensors for safety purposes. To achieve a fully unmanned robotic combine, a real‐time paddy field object detection method is necessary. Typically, all paddy field objects are detected individually using multiple algorithms and sensors, which significantly increases the complexity and cost of the detection process. In this study, the deep learning (DL) based semantic segmentation (SS) method was employed to detect all paddy field objects simultaneously using only an RGB camera. Considering the environment of the paddy field, a new SS model called “The Robotic Combine Network (TRCNet)” was specifically designed for the robotic combine. And four state‐of‐the‐art lightweight convolutional neural networks were applied as the backbones of the TRCNet. To achieve real‐time detection, TensorRT (NVIDIA) was utilized for speeding up the prediction process. All models were trained and evaluated using paddy field images captured during the robotic combine's harvesting process. The results showed that the TRCNet can successfully detect all paddy field objects. The mean intersection over union, and frames per second (FPS) of the best two SS models were 0.823, 47.48, and 0.834, 32.44, respectively. The FPS values were obtained after speed acceleration and tested with an image size of 640 × 480 pixels on an embedded processor (Jetson TX2), enabling real‐time object detection in paddy fields for the robotic combine.