Date of Award

June 2020

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Qinru Qiu


Computer Vision, Cyber Physical System (CPS), Deep Learning, Multi-agent System, Neural Network, Unmanned Aerial Systems (UAS)

Subject Categories



In recent years, the applications of Unmanned Aerial Systems (UAS) has become more and more popular. We envision that in the near future, the complicated and high density UAS traffic will impose significant burden to air traffic management. Lot of works focus on the application development of individual Small Unmanned Aerial Systems (sUAS) or sUAS management Policy, however, the study of the UAS cluster behaviors such as forecasting and managing of the UAS traffic has generally not been addressed. In order to address the above issue, there is an urgent need to investigate three research directions. The first direction is to develop a high fidelity simulator for the UAS cluster behavior evaluation. The second direction to study real time trajectory planning algorithms to mitigate the high dense UAS traffic. The last direction is to investigate techniques that rapidly and accurately forecast the UAS traffic pattern in the future. In this thesis, we elaborate these three research topics and present a universal paradigm to predict and manage the traffic for the large-scale unmanned aerial systems.

To enable the research in UAS traffic management and prediction, a Java based Multi-Agent Air Traffic and Resource Usage Simulation (MATRUS) framework is first developed. We use two types of UAS trajectories, Point-to-Point (P2P) and Man- hattan, as the case study to describe the capability of presented framework. Various communication and propagation models (i.e. log-distance-path loss) can be integrated with the framework to model the communication between UASs and base stations. The results show that MATRUS has the ability to evaluate different sUAS traffic management policies, and can provide insights on the relationships between air traf- fic and communication resource usage for further studies. Moreover, the framework can be extended to study the effect of sUAS Detect-and-Avoid (DAA) mechanisms, implement additional traffic management policies, and handle more complex traffic demands and geographical distributions.

Based on the MATRUS framework, we propose a Sparse Represented Temporal- Spatial (SRTS) UAS trajectory planning algorithm. The SRTS algorithm allows the sUAS to avoid static no-fly areas (i.e. static obstacles) or other areas that have congested air traffic or communication traffic. The core functionality of the routing algorithm supports the instant refresh of the in-flight environment making it appropri- ate for highly dynamic air traffic scenarios. The characterization of the routing time and memory usage demonstrate that the SRTS algorithm outperforms a traditional Temporal-Spatial routing algorithm.

The deep learning based approach has shown an outstanding success in many areas, we first investigated the possibility of applying the deep neural network in predicting the trajectory of a single vehicle in a given traffic scene. A new trajectory prediction model is developed, which allows information sharing among vehicles using a graph neural network. The prediction is based on the embedding feature, which is derived from multi-dimensional input sequences including the historical trajectory of target and neighboring vehicles, and their relative positions. Compared to other existing trajectory prediction methods, the proposed approach can reduce the pre- diction error by up to 50.00%. Then, we present a deep neural network model that extracts the features from both spatial and temporal domains to predict the UAS traffic density. In addition, a novel input representation of the future sUAS mission information is proposed. The pre-scheduled missions are categorized into 3 types according to their launching times. The results show that our presented model out- performs all of the baseline models. Meanwhile, the qualitative results demonstrate that our model can accurately predict the hot spot in the future traffic map.


Open Access

Included in

Engineering Commons