Control system for mobile agrobot based on neural network object detection
Вантажиться...
Дата
2025
Науковий керівник
Назва журналу
Номер ISSN
Назва тому
Видавець
КПІ ім. Ігоря Сікорського
Анотація
The paper addresses the urgent scientific and technical problem of automating crop care processes within the framework of the Precision Agriculture 4.0 paradigm. The primary objective of the study is to develop the architecture and investigate the efficiency of a cost-effective cyber-physical system (CPS) for autonomous crop monitoring and targeted physical destruction of weeds in real-time. This approach minimizes the reliance on chemical herbicides, thereby addressing the issues of weed resistance and environmental soil degradation. The authors propose and implement a two-level hierarchical control system for a mobile agricultural robot. The high-level computing layer is based on the Raspberry Pi 4 Model B single-board computer, which handles computer vision tasks and strategic path planning. The YOLOv8 Nano neural network architecture was selected and justified for semantic segmentation of vegetation cover. A set of optimization methods for Edge devices was applied, specifically model conversion to ONNX format and dynamic weight quantization to INT8 format. This reduced the model size to 6 MB and ensured stable inference on the CPU without hardware acceleration. The network training was conducted on a dataset of 10,000 images using a loss function that combines the IoU metric, binary cross-entropy, and Distribution Focal Loss. The low-level control is implemented on the ESP32 microcontroller (Dual Core architecture) running the FreeRTOS real-time operating system. Multi-threaded software was developed to separate communication tasks, inertial sensor (IMU) polling, and PWM signal generation. A discrete PID controller was implemented to stabilize the angular velocities of the differential drive platform wheels, compensating for errors caused by soil slippage. Inter-level communication is established via a UART interface (115200 baud) using a custom JSON-based protocol. A Direct Mapping method is proposed for manipulator control, eliminating the need for resource-intensive inverse kinematics calculations. Field test results confirmed the high efficiency of the system: a detection accuracy of mAP@0.5 at 92.4% was achieved with an average frame processing speed of 65–70 ms (14.5 FPS). The total latency in the control loop does not exceed 75 ms, and the platform positioning error is within ±2.5 cm. Energy monitoring indicated power consumption at the level of 18–22 W, providing up to 60 minutes of autonomous operation.
Опис
Ключові слова
agrobot, YOLOv8, Precision Agriculture 4.0, ESP32, Raspberry Pi, Computer Vision, PID controller, Boustrophedon, FreeRTOS, агроробот, YOLOv8, Precision Agriculture 4.0, ESP32, Raspberry Pi, Computer Vision, PID-регулятор, Boustrophedon, FreeRTOS
Бібліографічний опис
Tsompel, O. I. Control system for mobile agrobot based on neural network object detection / O. I. Tsompel, M. O. Bezuhlyi, Andrzej Dzierwa // Вісник КПІ. Серія Приладобудування : збірник наукових праць. – 2025. – Вип. 70(2). – С. 81-90. – Бібліогр.: 20 назв.