Date of Award

Spring 5-22-2021

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Qinru Qiu

Subject Categories

Computer Engineering | Engineering


Detection and control are two essential components in an intelligent system. This thesis investigates novel techniques in both areas with a focus on the applications of handwritten text recognition and UAV flight control. Recognizing handwritten texts is a challenging task due to many different writing styles and lack of clear boundary between adjacent characters. The difficulty is greatly increased if the detection algorithms is solely based on pattern matching without information of dynamics of handwriting trajectories. Motivated by the aforementioned challenges, this thesis first investigates the pattern recognition problem. We use offline handwritten texts recognition as a case study to explore the performance of a recurrent belief propagation model. We first develop a probabilistic inference network to post process the recognition results of deep Convolutional Neural Network (CNN) (e.g. LeNet) and collect individual characters to form words. The output of the inference network is a set of words and their probability. A series of post processing and improvement techniques are then introduced to further increase the recognition accuracy. We study the performance of proposed model through various comparisons. The results show that it significantly improves the accuracy by correcting deletion, insertion and replacement errors, which are the main sources of invalid candidate words.

Deep Reinforcement Learning (DRL) has widely been applied to control the autonomous systems because it provides solutions for various complex decision-making tasks that previously could not be solved solely with deep learning. To enable autonomous Unmanned Aerial Vehicles (UAV), this thesis presents a two-level trajectory planning framework for UAVs in an indoor environment. A sequence of waypoints is selected at the higher-level, which leads the UAV from its current position to the destination. At the lower-level, an optimal trajectory is generated analytically between each pair of adjacent waypoints. The goal of trajectory generation is to maintain the stability of the UAV, and the goal of the waypoints planning is to select waypoints with the lowest control thrust throughout the entire trip while avoiding collisions with obstacles. The entire framework is implemented using DRL, which learns the highly complicated and nonlinear interaction between those two levels, and the impact from the environment. Given the pre-planned trajectory, this thesis further presents an actor-critic reinforcement learning framework that realizes continuous trajectory control of the UAV through a set of desired waypoints. We construct a deep neural network and develop reinforcement learning for better trajectory tracking. In addition, Field Programmable Gate Arrays (FPGA) based hardware acceleration is designed for energy efficient real-time control.

If we are to integrate the trajectory planning model onto a UAV system for real-time on-board planning, a key challenge is how to deliver required performance under strict memory and computational constraints. Techniques that compress Deep Neural Network (DNN) models attract our attention because they allow optimized neural network models to be efficiently deployed on platforms with limited energy and storage capacity. However, conventional model compression techniques prune the DNN after it is fully trained, which is very time-consuming especially when the model is trained using DRL. To overcome the limitation, we present an early phase integrated neural network weight compression system for DRL based waypoints planning. By applying pruning at an early phase, the compression of the DRL model can be realized without significant overhead in training. By tightly integrating pruning and retraining at the early phase, we achieve a higher model compression rate, reduce more memory and computing complexity, and improve the success rate compared to the original work.


Open Access