Efficient Object Detection for Crop Monitoring in Precision Agriculture

Kerec, Jaša (author), Machidon, Octavian Mihai (mentor), Machidon, Alina - Luminita (comentor)

This thesis presents an automated approach to designing energy-efficient neural network architecture for wheat head detection in precision agriculture. By leveraging neural architecture search (NAS) on the YOLOv8n model, we developed optimized architecture tailored for deployment on edge devices such as the NVIDIA Jetson Nano and Raspberry Pi with OAK-D. Our best model reduced computational complexity by 37.0% GFLOPs and the number of parameters by 61.3% with negligible drop in detection accuracy (mAP@50). Furthermore, it achieved 28.1% improvement in FPS and 18.5% improvement in energy efficiency on NVIDIA Jetson Nano with ONNX runtime. Using TensorRT FP16 runtime it achieved 18.78% improvement in FPS and 39.34% improvement in energy efficiency. To asses the generalizability of the NAS approach we perform another search on plant seelding dataset yielding to 15.2% faster model (ONNX) with only negligible drop of 1.45% in accuracy.

• Language: English
• Keywords: neural networks, precision agriculture, object detection, neural architecture search
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2025
• PID: 20.500.12556/RUL-174663
• COBISS.SI-ID: 255394307
• Publication date in RUL: 08.10.2025

Secondary language

• Language: Slovenian
• Title: Učinkovita detekcija objektov za spremljanje pridelka v natančnem kmetijstvu

Abstract

V tej magistrski nalogi predstavljamo avtomatiziran pristop k zasnovi energetsko učinkovitih nevronskih arhitektur za detekcijo pšeničnih klasov v preciznem kmetijstvu. S pomočjo iskanja topologije nevronske mreže (angl. NAS) na modelu YOLOv8n smo razvili optimizirano arhitekturo, prilagojeno za izvajanje na robnih napravah, kot sta NVIDIA Jetson Nano in Raspberry Pi z OAK-D. Najboljši model je zmanjšal računsko zahtevnost za 37.0% GFLOPs in število parametrov za 61.3%, ob zanemarljivem padcu natančnosti. Poleg tega je na napravi NVIDIA Jetson Nano z okoljem ONNX runtime dosegel 28.1% večjo hitrost (FPS) in 18.5% boljšo energetsko učinkovitost na iteracijo. Z uporabo izvajalnega okolja TensorRT FP16 je dosegel za 18.78% večjo hitrost in 39.34% večjo energetsko učinkovitost. Da bi preverili generalizacijo pristopa NAS, smo izvedli še eno iskanje na podatkovni zbirki za detekcijo sadik, ki je prineslo 15.2% hitrejši model (ONNX) z zanemarljivim padcem 1.45% v natančnosti.

• Keywords: nevronske mreže, precizno kmetijstvo, detekcija objektov, iskanje topologije nevronske mreže