Defense Date

2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Electrical & Computer Engineering

First Advisor

Ruixin Niu

Abstract

Sensor fusion and tracking is the ability to bring together measurements from multiple sensors of the current and past time to estimate the current state of a system. The resulting state estimate is more accurate compared with the direct sensor measurement because it balances between the state prediction based on the assumed motion model and the noisy sensor measurement. Systems can then use the information provided by the sensor fusion and tracking process to support more-intelligent actions and achieve autonomy in a system like an autonomous vehicle. In the past, widely used sensor data are structured, which can be directly used in the tracking system, e.g., distance, temperature, acceleration, and force. The measurements' uncertainty can be estimated from experiments.

However, currently, a large number of unstructured data sources can be generated from sensors such as cameras and LiDAR sensors, which bring new challenges to the fusion and tracking system. The traditional algorithm cannot directly use these unstructured data, and it needs another method or process to “understand” them first. For example, if a system tries to track a particular person in a video sequence, it needs to understand where the person is in the first place. However, the traditional tracking method cannot finish such a task. The measurement model for unstructured data is usually difficult to construct. Deep learning techniques provide promising solutions to this type of problem. A deep learning method can learn and understand the unstructured data to accomplish tasks such as object detection in images, object localization in LiDAR point clouds, and driver behavior prediction from the current traffic conditions. Deep-learning architectures such as deep neural networks, deep belief networks, recurrent neural networks, and convolutional neural networks have been applied to fields including computer vision, speech recognition, natural language processing, audio recognition, social network filtering, and machine translation, where they have produced results comparable with human expert performance. How to incorporate information obtained via deep learning into our tracking system is one of the topics of this dissertation.

Another challenging task is using learning methods to improve a tracking filter's performance. In a tracking system, many manually tuned system parameters affect the tracking performance, e.g., the process noise covariance and measurement noise covariance in a Kalman Filter (KF). These parameters used to be estimated by running the tracking algorithm several times and selecting the one that gives the optimal performance. How to learn the system parameters automatically from data, and how to use machine learning techniques directly to provide useful information to the tracking systems are critical to the proposed tracking system.

The proposed research on the intelligent tracking system has two objectives. The first objective is to make a visual tracking filter smart enough to understand unstructured data sources. The second objective is to apply learning algorithms to improve a tracking filter's performance. The goal is to develop an intelligent tracking system that can understand the unstructured data and use the data to improve itself.

Rights

© The Author

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission

12-17-2021

Share

COinS