Efficient Object Detection and Segmentation Models for Drone Imagery

Drones offer unique aerial perspectives, making them valuable for various applications. However, robust object detection algorithms are crucial for extracting meaningful information from drone imagery. This research sees the YOLOv5 (You Only Look Once) model as a powerful solution for object detection within this setting. We look into the tradeoffs between performance and efficiency in the use of object detection and segmentation models, in particular, the various YOLOv5 models category under (nano, small, medium and large extralarge), various ResNet model types (18, 34, 50, 101), and in EfficientNet models type b0 up to b3. In the case of YOLOv5, there are smaller and larger versions depending on the use the user’s need of speed or accuracy respectively. ResNet models do not tend to increase accuracy with depth of the of the network but that is not without some gains being visible. Low and high-end EfficientNet models strike a fairly good compromise between accuracy and computational efficiency. We also test YOLOv5 models which include segmentation features, and it turns out that this increase in model size improves the segmentation capabilities. The work demonstrates that one cannot just select the best performance model in terms of accuracy accessible for the drone’s object detection task but also the balance between accuracy, speed and computational capabilities should be considered. To evaluate these models, we leverage the Roboflow dataset which is quite relevant for the drone imagery focused task. With object detection and image classification methods it becomes possible to recognize and process the data in drone images fast and correctly. This will allow for performance sustainability in systems operating in the real-time mode such as in monitoring and surveillance or search and rescue operations where speed and accuracy of target detection directly impacts the efficiency of the system.