Multi-UAV Trajectory Optimization and Deep Learning-based Imagery Analysis for a UAS-based Inventory Tracking Solution Youngjun Choi * , Maxime Martel † , Simon Briceno ‡ , Dimitri Mavris § Aerospace Systems Design Laboratory, School of Aerospace Engineering Georgia Institute of Technology, Atlanta, GA, 30332, USA This paper presents a multi-UAV trajectory optimization and an imagery analysis tech- nique based on Convolutional Neural Networks (CNN) for an inventory tracking solution using a UAS platform in a large warehouse or manufacturing environment. The current in- ventory tracking method is a manual and time-consuming process to scan all the inventory items. Its accuracy is not consistent depending on the complexity of the scanning environ- ment. To improve the scanning efficiency with respect to time and accuracy, this paper discusses a UAS-based inventory solution. In particular, this paper addresses two primary topics: multi-UAV trajectory optimization to scan inventory items and a multi-layer CNN architecture to identify a tag attached on the inventory item. To demonstrate the proposed multi-UAV trajectory optimization framework, numerical simulations are conducted in a representative inventory space. The proposed CNN-based imagery analysis framework is demonstrated on a flight experiment. I. Introduction In the last decade, the Unmanned Aerial System (UAS) has become more capable and advanced by new emerging technologies such as novel battery technologies and new sensing technologies. The use of a UAS can be beneficial because of its low-cost, highly agile platform and high-quality sensor system. Consequently, a UAS has been extended its application areas: aerial imaging, parcel delivery, crop-monitoring, and disaster monitoring. In particular, internal and external audits using a UAS platform in warehouse or manufacturing environments have gained attention because of the necessity of the novel structure of an audit evidence with respect to big data, and the improvement of the accuracy and speed of a traditional inventory audit process [2]. There are many challenges related to the internal/external audits, such as the identification of inventory items, their visual inspection, and counting them. Between the internal and external audits, this paper focuses on the internal audit process because the external audit process may be more challenging by FAA regulation. In particular, this paper deals with the inventory tracking problem to improve the current existing process in warehouse or manufacturing environments. A conventional approach for tracking inventory items is a manual scanning method using a barcode reader. Each item in a warehouse is equipped with a tag on which a barcode is printed. This barcode serves as a unique identifier for an individual item. Employee with a barcode scanner manually scans the barcode and identify the item. The scanned item is automatically stored in the inventory database to keep track of their locations. This manual scanning method has several drawbacks. First, it easily leads to human errors because of the highly repetitive nature of the task. Second, this scanning process is time-consuming * Research Engineer II, School of Aerospace Engineering, Aerospace Systems Design Laboratory, AIAA Member. † Graduate Research Associate, School of Aerospace Engineering, Aerospace Systems Design Laboratory ‡ Senior Research Engineer, School of Aerospace Engineering, Aerospace Systems Design Laboratory, and AIAA Senior Member. § S.P. Langley Distinguished Regents Professor, School of Aerospace Engineering, Aerospace Systems Design Laboratory, and AIAA Fellow. 1 of 14 American Institute of Aeronautics and Astronautics