Browsing by Author "Amayo, Paul"

Now showing 1 - 3 of 3
  • No Thumbnail Available
    Item
    Open Access
    A Low-cost autonomous tracking camera system for 3d marker-less motion capture of animals in the wild
    (2025) Vally, Amaan; Patel, Amir; Amayo, Paul
    The study of natural movement has long fascinated scientists, engineers and doctors. Today, motion capture research not only aids in medical diagnostics and rehabilitation but also enhances game and movie animations. Additionally, it contributes to the understanding of complex organic motions, informing the design of efficient, nature-mimicking robots. A large proportion of the research of human and animal motion capture relies on data captured using directional sensors with a limited field of view, such as RGB (red green blue) or RGB-D (red green blue-depth) cameras. Physical constraints limit the amount of data that can be collected with a single sensor (or set of sensors) since the subject is typically constrained to a specific capture area based on the sensor's field of view (FOV). This study focuses on the development of a camera-based system that can autonomously track a moving animal using rotating cameras to increase the amount of usable data that can be collected. In the pursuit of this objective, two systems were developed and tested. The first system consisted of a set of three cameras fixed to a rigid platform, with a camera on each end and the third midway between them. The platform was fixed to a brush-less DC (Direct Current) motor with the middle camera directly above the motor shaft. The second system consisted of an independent rotating camera fixed to the shaft of a brush-less DC motor. For both systems, the subject's position in the image frame of the camera mounted above the axis of rotation was determined using YOLO (You Only Look Once), a state-of-the-art object detection neural network. An extended Kalman filter (EKF) and full state feedback (FSF) controller were used to control the motor's position to keep the subject in the centre of the camera frame. DeepLabCut (DLC) was used to extract 2D key-points, and then a trajectory optimisation-based 3D pose estimation method called Full Trajectory Estimation (FTE) was used to reconstruct the 3D trajectories of the subject. Quantitative and qualitative experimental results are provided to validate the systems performance. Finally, this study concludes with recommendations for enhancing the system's performance, alongside proposed directions for future research and development in this field.
  • No Thumbnail Available
    Item
    Open Access
    Lightweight Mapping of Unstructured Environments
    (2024) Katsoulis, Michael; Amayo, Paul; Patel Amir
    We present a computationally, size and cost-lightweight monocular reconstruction pipeline that produces high-quality reconstructions in an unstructured agricultural environment consisting of orchards and vineyards. The pipeline has to be deployed and tested on ground vehicles equipped only with a single monocular camera sensor. Running on a CPU only, while achieving a sufficient resolution to identify individual plants without limitations on maximum path length. We show that a simple visual odometry system is capable of providing performance that is more accurate than GPS, with a relative transform error of 0.19m, without the need for techniques such as bundle adjustment or loop closure. Towards this contribution, we evaluate the impact of the choice of image feature on the accuracy of the visual odometry as well as the impact of the choice of disparity estimation method on the accuracy. Additionally, we show that state-of-the-art unsupervised monocular depth networks can outperform stereo techniques in terms of accuracy achieved when estimating the depth in an agricultural setting. We also show that lightweight pyramid-based methods are able to match the performance of deep monocular depth networks at the task of disparity estimation. The pipeline presented is optimised for application in agricultural environments and is lightweight in terms of size, weight, power and computational requirements. The pipeline functions using only a single camera and without any other sensors or sources of additional information making
  • No Thumbnail Available
    Item
    Open Access
    Localisation under Large Appearance Change
    (2024) Church, Matthew; Amayo, Paul
    Localisation is a foundational building block for more complex robot applications, and thus if low-cost localisation solutions can be found, the number of activities a robot can undertake will increase. However, appearance-based localisation systems in the past have required frequent traversals of the environment in order to sufficiently capture the change indicative of that environment. There are applications such as agriculture in which this frequent data collection is not appropriate. This thesis presents an appearance-based localisation system that combines generated and recorded data in the form of experience-localiser pairs combined to create an experience based network that can be used for localisation. The inclusion of generated data reduces the requirement for frequent data collection, provided an adequate generation model can be trained. The experience, which is a collection of images and transforms describing a traversal of an environment is the primary means through which this generation of data can influence the network. The images contained in the generated experiences were created from two parent experiences capturing two specific times of the day. The network trained learns a mapping from the two parent experiences creating intermediate domains that represent times between the parents, effectively filling in the gaps left by sparse data collection. While the performance of the generation network narrows the functional scope of the system, within that narrow scope, experiences generated from recorded outings outperform the recorded counterparts provided the parent does as well, such that an experience generated from a recording collected at 10:00 and made to mimic the conditions at 14:00 will outperform the recording collected at 14:00. Should a version be used such that all recorded experiences are utilized as a collective, the system outperforms that of a system making use of just recorded data