Applying computer vision to train unmanned aerial vehicles for recognizing basic geometric objects

Abstract

This study proposes a generalized methodology for training unmanned aerial vehicles (UAVs) to autonomously recognize objects of simple and common geometries in real-world environments. Specifically, a Convolutional Neural Network (CNN) model is integrated into the camera system of a DJI Tello drone to detect basic geometric objects, including rectangles, triangles, circles, and regular pentagons, which are selected to evaluate the model's recognition performance. A grayscale image dataset comprising objects of varying sizes and positions is automatically generated to optimize the data collection and model training process. The proposed CNN model, designed with a lightweight architecture to ensure real-time processing capability on the drone, is trained on this dataset and achieves approximately 100% accuracy on both the training and test sets. Subsequently, the model is integrated into the drone's camera system, and experimental results confirm its ability to perform accurate real-time object detection without overfitting. These findings demonstrate the effectiveness and practical potential of the proposed method for integration into intelligent drone systems.