The problem of image-based visual servoing (IBVS) of an aerial robot using deep-learning-based keypoint detection is addressed in this article. A monocular RGB camera mounted on the platform is utilized to collect the visual data. A convolutional neural network (CNN) is then employed to extract the features serving as the visual data for the servoing task. This paper contributes to the field by circumventing not only the challenge stemming from the need for man-made marker detection in conventional visual servoing techniques, but also enhancing the robustness against undesirable factors including occlusion, varying illumination, clutter, and background changes, thereby broadening the applicability of perception-guided motion control tasks in aerial robots. Additionally, extensive physics-based ROS Gazebo simulations are conducted to assess the effectiveness of this method, in contrast to many existing studies that rely solely on physics-less simulations. A demonstration video is available at https://youtu.be/Dd2Her8Ly-E.

This paper introduces an innovative keypoint detection technique based on Convolutional Neural Networks (CNNs) to enhance the performance of existing Deep Visual Servoing (DVS) models. To validate the convergence of the Image-Based Visual Servoing (IBVS) algorithm, real-world experiments utilizing fiducial markers for feature detection are conducted before designing the CNN-based feature detector. To address the limitations of fiducial markers, the novel feature detector focuses on extracting keypoints that represent the corners of a more realistic object compared to fiducial markers. A dataset is generated from sample data captured by the camera mounted on the robot end-effector while the robot operates randomly in the task space. The samples are automatically labeled, and the dataset size is increased by flipping and rotation. The CNN model is developed by modifying the VGG-19 pre-trained on the ImageNet dataset. While the weights in the base model remain fixed, the fully connected layer's weights are updated to minimize the mean absolute error, defined based on the deviation of predictions from the real pixel coordinates of the corners. The model undergoes two modifications: replacing max-pooling with average-pooling in the base model and implementing an adaptive learning rate that decreases during epochs. These changes lead to a 50 percent reduction in validation loss. Finally, the trained model's reliability is assessed through k-fold cross-validation.