Quantized Coconut Detection Models with Edge Devices

Abstract

Coconut is a multipurpose fruit with high economic value and since it is unique to the landscape of Kerala, it plays an important role in the economy of the state. Skilled labour is one of the key components in coconut farming and lack of its availability can hurt its business. Even if this requirement is met, currently practiced traditional methods for plucking the fruit requires the labour to climb the tree which involves a huge risk factor given the height of the tree they have to scale. There are tools that assist in the climb but they can only reduce the risk factor by a small margin. Robotic harvesting is one of the key solutions to the aforementioned problem as it has the ability to perform accurate coconut plucking since it relies on cutting edge object detection modules, it can provide deep insights into the quality of coconuts to be yielded and also excel at working in remote conditions. The primary aim of this paper is to cover the development of a fast as well as accurate perception module for detection of coconuts, which will serve as a strong foundation for any robotic implementation. In this study we try to explore and compare multiple deep learning based object detection frameworks such as Single Shot Detector and YOLO for efficient and accurate deployment on various edge devices like Raspberry Pi and Nvidia jetson nano by using state of the art methods such as quantization aware training, inference accelerators, multiple augmentation strategies (cutmix, mosaic) for best results. We have also curated a novel, manually annotated dataset of drone based coconut videos (effective/usable content of 30 minutes) in order to capture the naturally setting of coconuts i.e. the true distribution of image data containing background noises, occlusion, shadow as well as natural lighting conditions. The peak performance achieved in our study was a frame rate of 12.7 with a mean average precision of 0.4 by using a tiny YOLOv4 on an Nvidia Jetson Nano.