Robust Object Recognition and Tracking with Drones

Abstract: The Skara Skyddsängel project explores an innovative method of providing illumination for cyclists along a 20km unlit bike lane using drones. Current GNSS approach performs generally well but further improvements are need for better robustness. Consequently, this thesis project is raised to seek a robust solution in the field of computer vision. YOLOv7 is one of the most advanced object detection algorithms. Considering that this will be applied to drones, we opted for the compact version, YOLOv7-tiny. To enhance the performance of the algorithm, we created a custom dataset and expanded it through data augmentation techniques. After our comparative experiments, the model reached a favourable mean average precision (mAP) of 98.88%. Evaluation on unseen videos shows that the trained model can effectively meet the specific requirements of this unique scenario.