Capturing and Classifying Complexity for Pedestrian Environments

Abstract

Environments in which people navigate on foot are referred to as pedestrian environments. Developing navigation strategies for pedestrian environments is challenging due to the diversity of scenes to tackle. Most solutions focus on specialized strategies for more narrow tasks. These methods perform well in their closed scenarios but could struggle when applied to others. There are no solutions yet that have achieved generality for all scenes in pedestrian environments. This thesis explores the option of selecting the optimal navigation strategy from the current state of the environment, referred to as Real-Time Strategy Selection (RTNSS) as a solution. As this is a relatively unexplored solution, the research conducted in this thesis attempts to lay the needed groundwork for experiments in a highly simplified environment featuring crowds of pedestrians in an otherwise empty hallway. For these purposes, a novel annotation algorithm is built by analyzing sensor signals from a wheelchair operator and annotating images as Calm or Busy. This algorithm is then used to construct a dataset, named the Hallwayset which allows for the training of models for future experiments. The annotation algorithm was tested on a series of indoor scenarios and was shown to be capable of matching manual annotations. The Hallwayset was too small for proper model evaluations but could be expanded using the annotation algorithm. A pre-trained ResNet18 model was trained to classify images as Busy or Calm but made unstable and overly cautious predictions. This could be solved by adding context to the image. The conducted research laid a foundation on which experiments can be conducted to determine the quality of navigation strategy selection. The resources provided allow for simple model training, as well as rapid collection and annotation of data. The learnings provide valuable insight into how to construct a model for real-time navigation strategy selection. However, modifications to the resources provided in this thesis need to be made for future experiments.