Logistics Hub Surveillance: Optimizing YOLOv3 Training for AI-Powered Drone Systems

Background: Integrating artificial intelligence in unmanned aerial vehicle systems may enhance the surveillance process of outdoor expansive areas, which are typical in logistics facilities. In this work, we propose methods to optimize the training of such high-performing systems. Methods: Specifically, we propose a novel approach to tune the training hyperparameters of the YOLOv3 model to improve high-altitude object detection. Typically, the tuning process requires significant computational effort to train the model under numerous combinations of hyperparameters. To address this challenge, the proposed approach systematically searches the hyperparameter space while reducing computational requirements. The latter is achieved by estimating model performance from early terminating training sessions. Results: The results reveal the value of systematic hyperparameter tuning; indicatively, model performance varied more than 13% in terms of mean average precision (mAP), depending on the hyperparameter setting. Also, the early training termination method saved over 90% of training time. Conclusions: The proposed method for searching the hyperparameter space, coupled with early estimation of model performance, supports the development of highly efficient models for UAV-based surveillance of logistics facilities. The proposed approach also identifies the effects of hyperparameters and their interactions on model performance.