Award Date

8-1-2021

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

First Committee Member

Brendan Morris

Second Committee Member

Venkatesan Muthukumar

Third Committee Member

Emma Regentova

Fourth Committee Member

Mingon Kang

Number of Pages

67

Abstract

In this dissertation we develop different methods for forecasting pedestrian trajectories. Complete understanding of pedestrian motion is essential for autonomous agents and social robots to make realistic and safe decisions. Current trajectory prediction methods rely on incorporating historic motion, scene features and social interaction to model pedestrian behaviors. Our focus is to accurately understand scene semantics to better forecast trajectories. In order to do so, we leverage semantic segmentation to encode static scene features such as walkable paths, entry/exits, static obstacles etc. We further evaluate the effectiveness of using semantic maps on different datasets and compare its performance with already existing scene encoding approaches based on CNN/VGG16 architectures. Our experiments and results have shown significant improvement which validated our hypothesis regarding the efficiency of using fully segmented maps.

We also focus on better learning inter-pedestrian behavior. When pedestrians walk they tend to avoid collision with other pedestrians. Previously pedestrian-pedestrian interactions were modeled using social pooling techniques where a grid like structure with certain neighborhood is considered and then this information is passed through some neural network to capture social information. In our work we build a robust architecture to incorporate these social interactions via Graph Convolutional Neural Networks (GCN) and Transformers. GCN first, extracts Spatio-Temporal (ST) representations of pedestrians. Scene information is included through fully segmented output map and then an attention framework is applied on ST representations and scene features through transformers to predict future trajectories. Our experimental evaluation shows that such modeling technique shows significant decrease in prediction error compared to other grid pooling methods.

In our last work we propose a new frame work to forecast long term trajectories based on Inverse Reinforcement Learning. Simply learning input/output mapping of sequences for longer horizons tend to accumulate errors, therefore we come up with an approach which tends to learn an optimal policy for longer horizons via learning intermediate short term trajectories.

Keywords

Autonomous Driving; Computer Vision; Deep Learning; Machine Learning; Pedestrian Trajectory Prediction

Disciplines

Artificial Intelligence and Robotics | Computer Engineering | Computer Sciences | Electrical and Computer Engineering

File Format

pdf

File Size

8800 KB

Degree Grantor

University of Nevada, Las Vegas

Language

English

Rights

IN COPYRIGHT. For more information about this rights statement, please visit http://rightsstatements.org/vocab/InC/1.0/

Download


Share

COinS