Comparative Evaluation of Camera Modules for Real-time Deep Learning-Based Detection of Tomato Plants and Weeds on a Raspberry Pi Platform
Apoorva Sharma *
Department of Farm Machinery and Power Engineering, College of Technology, G. B. Pant University of Agriculture and Technology, Pantnagar, Uttarakhand, India.
Arun Kumar
Department of Farm Machinery and Power Engineering, College of Technology, G. B. Pant University of Agriculture and Technology, Pantnagar, Uttarakhand, India.
Hemant Kumar Sharma
Department of Farm Machinery and Power Engineering, College of Technology, G. B. Pant University of Agriculture and Technology, Pantnagar, Uttarakhand, India.
*Author to whom correspondence should be addressed.
Abstract
Aims: To evaluate and compare three camera modules, namely a USB webcam, a smartphone camera, and a Raspberry Pi Camera Module, for real-time detection of tomato plants and weeds, in order to identify the most suitable imaging sensor for deployment in a precision spraying system.
Study Design: Comparative experimental evaluation of camera hardware under static and simulated dynamic field conditions.
Place and Duration of Study: Department of Farm Machinery and Power Engineering, G. B. Pant University of Agriculture and Technology, Pantnagar, Uttarakhand, India, between July and September 2024.
Methodology: A You Only Look Once (YOLO)11n object detection model, trained to identify tomato plants and weeds, was deployed on a Raspberry Pi 4B for real-time inference. Three imaging sensors, a 2-megapixel (MP) USB webcam, a 64 MP smartphone camera (POCO M4 Pro) interfaced via the DroidCam application over Wi-Fi, and a 5 MP Raspberry Pi Camera Module interfaced through the Camera Serial Interface (CSI), were evaluated under static conditions and under conveyor-simulated dynamic conditions at 1 km/h. The modules were compared on resolution, interface type, integration complexity, image quality, detection accuracy, processing latency, and power requirement.
Results: The USB webcam exhibited poor image quality and a processing latency of 1250–1365 milliseconds (ms). The smartphone camera achieved the highest image quality and detection accuracy but showed a latency of approximately 2,000–3,000 ms due to Wi-Fi-based streaming. The Raspberry Pi Camera Module recorded the lowest latency (700–900 ms), moderate-to-high detection accuracy, and required no external power source, giving it the most favourable overall balance among the three sensors.
Conclusion: Data transmission pathway, rather than sensor resolution alone, was the decisive factor determining real-time deployability. The Raspberry Pi Camera Module was accordingly selected for the final precision spraying prototype, and these findings offer practical guidance for imaging sensor selection in edge-deployed, real-time precision agriculture systems.
Keywords: YOLO11n, Raspberry Pi, real-time detection, precision agriculture, latency