PRACTICAL COMPARISON OF NEURAL NETWORK MODELS IN SOLVING THE PROBLEM OF DETECTING OBJECTS IN IMAGES COMING FROM UAVS

The paper compares two neural network models to find the best ones in the tasks of detecting incoming objects from the television system of an unmanned aerial vehicle (UAV). Two open-source models were used, namely YOLOv8n and SSD512 (ResNet). Further training and comparison of models obtained on the VisDrone2019-DET dataset, designed for training models for working with images coming from UAVs. The models were compared based on mAP indicators - the average accuracy model for determining object classes, the accuracy of their work and completely correct answers. In addition, the speed of work was assessed based on the number of calculated frames per second (FPS). It is shown that YOLOv8n outperforms SSD512 (ResNet) in all performance indicators and is best suited for use on UAVs.