W

Retaining the balance between high productivity and quality while efficiently managing available resources is one of the biggest challenges facing the agricultural sector. As a possible solution to overcome these challenges and obstacles, precision agriculture has emerged in recent years. Among the promising solutions that precision agriculture offers, is the use of edge computing devices for monitoring and acquiring data in the rural environment, processing information locally and in real time. Computer Vision and Artificial Intelligence, more specifically referred to as Deep Learning, have also being applied recently in agriculture for different tasks such as image classification, object detection and semantic segmentation. However, there is a challenge and limitation in transferring this technology to more affordable platforms to process large quantities of data. Therefore, in this work, we explored the use of Computer Vision and Deep Learning applied to the object detection task in edge devices, specifically the Raspberry PI 4 platform, without hardware acceleration. We decided to apply this methodology for weed detection, since weeds are currently one of the pests that result in greatest loss of productivity in agriculture, and also have rapidly developed resistance to commercial herbicides. Also, in order to evaluate the performance gain for real-time weed detection on the Raspberry Pi platform, quantizaion of the deep neural network architectures using TensorFlow Lite was tested. The experimental results suggests that the proposed methodology is functional, and it was achieve real time weed detection on Raspberry Pi 4 platform, making it possible to reproduce the experimental results on similar edge devices.