City Research Online

Abstract

Robotic applications are continuously expanding into every aspect of human livelihood, it becomes paramount to leverage this trend for precision agriculture. The agricultural sector despite being an important sector for human is slowly evolving in terms of technology. Crude and manual processes which are conventionally used for agriculture have severe economic and social impacts. The inefficiencies and less productiveness of these methods results to food wastage amidst food shortage, inconsistencies, time consumption, higher labour expenses, and low yield. The world will benefit from automating the processes in agriculture. In bid of addressing such, it becomes necessary to build on existing platforms and develop intelligent autonomous vehicles for precision agriculture. This should include development of intelligent drones for precision agriculture, development of intelligent ground robots for precision agriculture, and other systems working cooperatively. To achieve this, we leverage on Artificial Intelligence (AI) and mathematical methods to impact sufficient intelligence on robotic platforms to make them suitable for precision agriculture.

This thesis explores the capabilities of AI for weed classification and detection, weed relative position estimation, fruit 6D pose estimation and virtual reality for teleoperated systems in fruit picking. Infestation of weeds diminishes the yield of crops in agriculture. Deep learning is becoming a more popular approach for identifying weeds on farmlands. However, precision agriculture requires that the object of interest (weed) is precisely classified and detected to facilitate removal or spraying. An approach for this is presented and involves cascading a classification network (ResNet-50) with a detection network (YOLO) for weed classification and detection which we termed Fused-YOLO. Thus, weeds can precisely be located and classified (type) within an image frame.

Inspired by the precision of this detection model, the work extends to presenting a novel monocular vision-based approach for drones to detect multiple types of weeds and estimate their positions autonomously for precision agriculture applications. A drone is subjected to an elliptical trajectory while acquiring images from an onboard monecular camera. The images are fed to the fused-YOLO model in real-time. The centre of the detection bounding boxes is leveraged to be the centre of the detected object of interest (weeds). The centre pixels are extracted and converted into world coordinates forming azimuth and elevation angles from the target to the UAV and are effectively used in an estimation scheme that adopts the Unscented Kalman Filteration to estimate the exact relative positions of the weeds. The robustness of this algorithm allows for both indoor and outdoor implementation while achieving a competitive result with affordable off-the-shelf sensors.

Artificial intelligence for autonomous 6D pose estimation has valuable contributions to agricultural practices rallying around fruit picking, harvesting, remote operations and other contact-related applications. Conventionally, Convolutional Neural Networks (CNNs) based approaches are adopted for pose estimation. However, precision agriculture applications are demanding on higher accuracy at lower computational costs for real-time applications. Motivated by this, a novel architecture called Transpose is proposed based on transformers. TransPose is an improved Transformer-based 6D pose estimation with a depth refinement. More modalities often result in higher accuracy at the expense of computational cost. TransPose takes in a single RGB image as input without extra modality. However, an innovative light-weight depth estimation network architecture is incorporated into the model to estimate depth from an RGB image using a feature pyramid with an up-sampling method. A transformer model having proven to be efficient, regress the 6D pose directly and also outputs object patches. The depth and the patches are utilised to further refine the regressed 6D pose. The performance of the model is extensively assessed and compared with state-of-the-art methods. As part of this research, a first-ever fruit-oriented 6D pose dataset was acquired.

Lastly, a seamless teleoperation pipeline that interfaces virtual reality with robots for precision agriculture tasks is proposed to pave the way for virtual agriculture. This utilises the Transpose model to estimate the 6D pose of a fruit and render it in a virtual reality environment. A robotic manipulator is which is then controlled from within the virtual reality environment to pick/harvest the fruit while being guided by the Transpose AI model. The robustness of the pipeline is tested over simulation and real-time implementation with a physical robotic manipulator is also investigated.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA76 Computer software S Agriculture > S Agriculture (General)
Departments:	School of Science & Technology > Engineering School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads