Features Exploitation of YOLOv5-Based Freeze Backbone for Performance Improvement of UAV Object Detection

Object detection in unmanned aerial vehicles (UAV) has steadily developed over time. Improving the performance of UAV-based object detection has proven challenging for researchers because of several problems, such as data imbalance, scale variants, prediction accuracy, and memory and computation limitations for real-time applications. In this study, feature exploitation using the YOLOv5 algorithm is carried out to improve the performance of UAV-based object detection. The best-performing model leverages a frozen backbone, modifies concatenation layers, and adds a head to the algorithm for improved performance. This approach improved performance and achieved a mAP@0.5 value of 22.4%. Based on the results of the mAP, the exploitation of features affected the performance of UAV-based object detection.