Real-time counting of strawberry using cost-effective embedded GPU and YOLOv4-tiny

Abstract. Recent years have seen an increase in demand for strawberries, which necessitates automation in the relevant agricultural processes. Multiple object detection models have been proposed previously in order to automate agricultural processes, through applications such as fruit and disease detection. However, in integrating these improvements, it is essential that the deployment costs and weight are not increased significantly. This paper presents an analysis of the performance and accuracy of YOLOv4 and YOLOv4-tiny on real-time strawberry detection when inference is carried out on an embedded GPU device such as an NVIDIA Jetson Nano. Three frameworks (Darknet, TensorRT, and TensorFlow Lite) and three different resolutions (416x416 pixels, 480x480 pixels, and 640x640 pixels) were compared to obtain an optimal mAP and inference speed trade-off. The proposed setup is cost-effective and lightweight, which is good for small, unmanned ground vehicles and/or drones, and can achieve an accuracy of 91.95% at an FPS rate of 14.6, which makes it a viable option for deployment in strawberry fields.