Landmark selection for route instructions: At which corner of an intersection is the preferred landmark located?

Hamburger, Kai; Röser, Florian; Knauff, Markus

doi:10.3389/fcomp.2022.1044151

ORIGINAL RESEARCH article

Front. Comput. Sci., 10 November 2022

Sec. Human-Media Interaction

Volume 4 - 2022 | https://doi.org/10.3389/fcomp.2022.1044151

Landmark selection for route instructions: At which corner of an intersection is the preferred landmark located?

$\nKai Hamburger$ Kai Hamburger¹^*

Florian Röser²

Markus Knauff¹

¹Department of Psychology, Experimental Psychology and Cognitive Science, Justus Liebig University, Giessen, Germany
²Department of Social Sciences, University of Applied Sciences, Darmstadt, Germany

Cognitive studies showed that good landmarks–salient objects in the environment–make it easier for recipients of route instructions to find their way to the destination. Adding landmarks to route instructions also improves mobile navigation systems for pedestrians. But, which landmarks do people consider most helpful when giving route instructions? Four experiments explored this question. In the first experiment, the environment, including the route and landmarks, was presented on a map. The landmarks were located at the four corners of a right-angled intersection. Participants had to select those landmark-based route instructions they considered most helpful. In all other experiments, the environment was presented from an egocentric perspective, either in a video or as a sequence of pictures of intersections. Participants had to select those landmarks they would use in a route instruction. All landmarks had the same visual and semantic salience. The positions of the participants at the intersection were varied. Results show that participants consistently selected landmarks at the side of the road into which they had to turn. Moreover, the participants' position at the intersection affected whether they selected landmarks before or behind the decision point. These results have consequences for human spatial cognition research and for the automatic selection of landmarks in mobile pedestrian navigation systems.

Introduction

Landmarks are one of the central concepts in spatial cognition research (Siegel and White, 1975; Montello, 2017; for a recent review see Yesiltepe et al., 2021). Most broadly, they are defined as easily recognizable objects that stand out from the environment and thus help wayfinders to determine their location or to describe the route to another person (Denis et al., 1999; Sorrows and Hirtle, 1999). People can select almost everything as a landmark from skyscrapers to small houses, all kinds of human-made objects such as sculptures in the public space. They can also use natural objects such as trees at an intersection, they can, of course, use signage, and they can even use the specific topography of the environment, for instance, when they say “go over the hill” (Tom and Denis, 2004; Tom and Tversky, 2012).

Landmarks are also important for mobile pedestrian navigation systems (Ross et al., 2004). It has also been demonstrated that both, pedestrians as well as car drivers profit from landmark inclusion in navigation instructions (e.g., Burnett, 2000; May and Ross, 2006; Wunderlich and Gramann, 2021). However, slight differences in the type of landmark information given might occur, for example, due to the speed with which car drivers and pedestrians move along their ways (i.e., the car driver has less time for her perceptions, route decisions, and actions, while pedestrians can focus on many more details in the environment, since they have much more time, and if necessary, can simply turn around, etc.). While car drivers primarily need access to the geometry of intersections and turn-by-turn instructions such as “In 300 meters, turn left into State Street,” pedestrians also need information about landmarks, such as in “At the City hall, turn right” or “Behind the church, turn left.” In the last decades, such landmark-based instructions became increasingly important for the development of pedestrian navigation systems (Millonig and Schechtner, 2005, 2007; Ohm et al., 2015). However, it is still challenging to develop techniques for the automatic extraction of suitable objects for pedestrian navigation instructions from the massive amount of available geospatial data from the internet and social media (Raubal and Winter, 2002; Elias, 2003; Winter et al., 2008; Selvi et al., 2012; Rousell et al., 2015). The present paper is not concerned with the conceptualizations and technical challenges in the development of pedestrian navigation systems (e.g., Hansen et al., 2006; Klippel et al., 2012). Our aim is rather to take a cognitive perspective on some of these challenges.

Developers of mobile navigation systems for pedestrians devised several methods for effective selection of landmarks for navigational instructions. We can distinguish two different lines of research: One is pure engineering and does not care much for cognitive considerations (e.g., Mohinder et al., 2001). The other branch of research aims to develop human-centered methods to help users to comfortably navigate through unknown environments. Such approaches often devise cognitively inspired methods for extracting landmarks from new sources, in particular social media (Quesnot and Roche, 2014; Zhu and Karimi, 2015) or OpenStreetMap (OSM) that are open, globally accessible, and provide enormous amounts of symbolic or pictorial information about potential landmarks. For example, Rousell et al. (2015) developed methods for extracting landmarks from OSM based on distance and estimated visibility. Rousell and Zipf (2017) devised algorithms for the extraction, weighing, and selection of landmarks based on their suitability for the generation of landmark-based navigation instructions for pedestrian routes. Raubal and Winter (2002) and Klippel and Winter (2005) proposed measures to formally specify the salience of landmarks for route instructions.

Cognitive scientists have identified several characteristics that contribute to the salience of potential landmark objects. Please note that not just objects can represent landmark information but also other sensory information (e.g., auditory or olfactory; e.g., Karimpur and Hamburger, 2016; Hamburger and Knauff, 2019) and regional-like features such as a park or lake as well as line features such as rivers or rail tracks can be considered as well (e.g., Anacta et al., 2017; Schwering et al., 2017; Löwen et al., 2019). In general, the salience of an object (or other type of information) is the quality by which it stands out from its surrounding. Psychologists see salience detection as a key attentional mechanism that enables people to deal with their limited perceptual and cognitive resources in order to act efficiently in their environment (Goldstein, 2015). For landmark selection, we can distinguish three types of salience: visual, semantic, and spatial salience (e.g., Caduff and Timpf, 2008; Nuhn and Timpf, 2017a,b). For a better and more comprehensive understanding, we want to briefly introduce all of them, even though the focus of the current work will be on the spatial/structural aspects of the environment. Visual salience is given, for instance, if a building with a unique shape or color stands out from its neighbors. It is related to findings from visual perception and attention research showing that objects that stand out from their surroundings quickly reach the focus of attention (e.g., Treisman and Gelade, 1980; Wang and Theeuwes, 2020). Researchers have extensively investigated how these factors affect landmark selection (e.g., Appleyard, 1969; Itti and Koch, 2001; Jin et al., 2004; Röser et al., 2013; Butz, 2014).

While the importance of visual features in landmark-based wayfinding has repeatedly been demonstrated (for review see Epstein and Vass, 2014), other research challenges this overarching importance of visual aspects in comparison to other sensory modalities, i.e., audition. For instance, Hamburger and Röser (2014) found that acoustic landmarks can be equally helpful for wayfinding. More generally, Knauff and Johnson-Laird (2002) could show that too much visual information can even hinder cognitive processes, such as spatial reasoning or problem-solving. They can hinder the construction of mental models, or, more generally, the generation of a mental representation from a given verbal (route) description (Knauff, 2013). Given that, we tried to keep the visual salience of landmark objects as constant as possible for the purpose of our study.

Semantic salience is obtained if the building has a particular meaning or function, like a police station in comparison to a regular building without any special relevance (Sorrows and Hirtle, 1999; Nothegger et al., 2004; Caduff and Timpf, 2008). While the visual salience of landmark objects is largely driven by bottom-up processes of perception and attention, semantic salience is based on top-down processes, in which prior knowledge is retrieved from long-term memory. Such retrieval processes can rely on semantic memory, i.e., on our world knowledge, ideas, concepts, beliefs, attitudes, and everything that we have accumulated throughout our lives that helps us to make sense of the world. For example, the knowledge that red buildings are often fire stations is stored in semantic memory (for an overview, see Anderson, 2000). Semantic salience can also rely on episodic memory, i.e., on our biographical knowledge about events, which are represented including the temporal and spatial context. For instance, an instruction-giver might say “Turn right at the bar where we were last Saturday.” Such personal episodic knowledge is, of course, difficult to use in navigation systems. Yet, it might be usable at least in rudimentary form, e.g., when locations where the user has been before are re-used in a new route description. In order to control for semantic salience, we tried to keep it as low as possible and constant (please see our initial results for the chosen material in Table 1) in order to systematically investigate the third type of salience.

TABLE 1

Table 1. Distribution of the visual object properties in all experiments.

This third kind of salience, spatial salience, is the topic of this paper. Some researchers refer to these location-related aspects of landmarks as structural salience. They often emphasize how important this kind of salience is for human wayfinding (Lovelace et al., 1999; Sorrows and Hirtle, 1999; Steck and Mallot, 2000; Caduff and Timpf, 2008; Hamburger and Knauff, 2011; Röser et al., 2012). We agree with this position but think that the term “structural salience” is less clear than the term “spatial salience,” which better expresses the fact that this kind of salience is related to the spatial location of landmarks in the environment. As described in the next section, the goal for the present work is to develop a better cognitive understanding of how the spatial salience of different landmark objects affects people's landmark selection. Are landmarks at a certain position of an intersection more salient than others? Do people prefer landmarks before or behind the decision point, at the same side of the street or on the opposite side?

Aims and methodology of the present experiments

Let us illustrate the research question with an easy example: Consider Figure 1 and imagine a tourist pedestrian asking you for directions to the train station. You know that she has to turn left at this intersection. How will you provide her with this information? Alternatively, you can also imagine a person who uses a navigation App on her smart phone. Which landmark should the system use in the verbal instructions?

FIGURE 1

Figure 1. Example of an intersection with four landmarks. Which of these landmarks would you use in a route description if the person has to turn left to reach her destination?

Here are a few possibilities that could be helpful for the person to find her way to the train station:

Turn left at the intersection.

Turn left behind the hospital.

Turn left before you pass the police station.

Turn left directly before the gas station.

Turn left where the church is on the right.

Of course, many other possibilities are conceivable. But which alternative do people choose? Obviously, alternative 1 is the simplest description and actually sufficient to tell a pedestrian in which direction to turn at the intersection. However, you may consider other landmarks as more helpful when you generate your wayfinding instruction. For example, you might refer to certain landmarks just based on their position in the environment. The goal of the following four experiments therefore is to find out which landmarks at which position humans consider most helpful in route instructions. Note that this is not the same as the question which of the landmarks would actually be optimal or most helpful for the recipient of the wayfinding instruction. This is a different question, to which we return in the General Discussion.

In the following sections, we report four experiments in which human participants were asked to select those landmarks at an intersection that they considered most helpful for a route instruction in this situation and setup. To avoid effects of visual and semantic features, we used colored geometrical shapes as landmarks. We also varied the position of the participants at the crossroad. This is an important variation to previous experiments in this domain (for review see (Röser, 2015)) because in daily life we also often stand at different positions of an intersection, which can make some landmarks more or less salient and helpful than others. All experimental materials were generated from the SQUARELAND environment, which we have already used in several of our previous experiments (Röser et al., 2011; Hamburger and Röser, 2014; Röser, 2015; Karimpur and Hamburger, 2016; Hamburger and Knauff, 2019), and which is now also used in other labs (Albrecht and von Stuelpnagel, 2018). SQUARELAND basically consists of a 10 x 10 block raster with orthogonal intersections. The blocks can be flexibly adapted with respect to size, surface structure, and so on (Hamburger and Knauff, 2011). In the first experiment, which was largely explorative, the environment, including the route and landmarks, was presented on a map. The landmarks were located at the four corners of a right-angled intersection. Participants had to select those landmark-based route instructions they considered most helpful. In all other experiments, the environment was presented from an egocentric perspective, either in a video or as a sequence of pictures of intersections. Participants had to select those landmarks they would use in a route instruction. All experiments used simple colored geometrical figures as landmarks to avoid uncontrolled effects of visual and semantic salience. In Experiments 1 and 2, we combined four shapes (square, trapezoid, triangle, and circle) with four colors (blue, green, red, and yellow), resulting in 16 landmarks. In Experiments 3 and 4, we used four colored circles. These materials were already used and evaluated in a pilot study (Table 1) and are visualized in Figure 2. A statistical analysis for the present experiments showed that the visual object properties had no significant effect on participants' landmark preferences: squares, trapezoids, triangles, circles, as well as blue, green, red, and yellow objects were selected equally often (Table 1). Since these objects also did not vary in semantic salience, we can attribute participants' landmark selections in the following experiments just to their locations at the intersection.

FIGURE 2

Figure 2. Materials and setups used in the experiments.

For the statistical analysis of all experiments, we used Kendall's W to assess participants' agreement in choosing landmarks. Kendall's W is a non-parametric test that should be used when the assumptions of parametric tests are violated (Siegel and Castellan, 1988; Hollander et al., 2013). It allows to determine whether there is statistically significant agreement across participants on which of the four landmark objects they choose at the intersection. Kendall's W ranges from 0 (no agreement) to 1 (complete agreement). Typically, the following interpretation guidelines are used: 0.1 – < 0.3 small effect, 0.3 – < 0.5 moderate effect, and > 0.5 large effect (Hollander et al., 2013). When multiple tests were computed, the p values were corrected with the Bonferroni method to avoid α-error accumulation. Using just Kendall's W might be a relatively weak statistical method and could be accompanied by other statistics, e.g., mixed-models statistics to analyze both fixed and random effects. Unfortunately, this is not possible due to a technical problem leading to a data format that did not allow for more sophisticated statistical analyses. Of course, on the one hand, this weakens our general findings and therefore our results must be interpreted with caution. On the other hand, most differences are clearly visible in the descriptive data and we did not perform any p-hacking to make minor differences statistically significant.

Experiment 1: Map presentation

Our first experiment was exploratory. It tested whether people select route instructions with landmark objects at certain positions of the intersection more often than ones with landmarks at other positions. At each intersection, four landmarks were presented and therefore each of them would have a 25 percent chance of being selected if participants did not prefer a particular landmark location. Yet, we predicted that some landmarks should be selected more often than others. This experiment was an online experiment and the environment was presented from a bird's-eye perspective on the map.

Participants

Twenty-six students (18 females, 8 males) participated in the experiment. The mean age was 22.9 years, with a range of 19–38 years. They were recruited via the circular email system of the university (since in the online experiment there was no personal contact and data were collected anonymously, it was possible to participate in one of the other experiments; but only in one of them in order to avoid any effects based on previous experience in the egocentric perspective). They all provided informed written consent and participated voluntarily. If required, they could receive course credits for participation. All experiments were reviewed and approved by the German Psychological Society (DGPs; MK3010200DGPS).

Materials and design

Each participant had to learn one route through the grid-like SQUARELAND environment with 7 × 4 orthogonal streets, making a total of 28 intersections. The route consisted of 8 left and 8 right turns at 16 intersections. The sequence of left and right turns was pseudo-random, with the limitation that walking in circles was impossible (i.e., three successive turns to either left or right) and the route remained within the maze grid. For each participant an individual sequence of intersections was provided in order to control for sequence or position effects. At each intersection, four landmarks were shown and the participants had to select one of four route instructions, each referring to one of these landmarks. The landmarks were the 16 colored geometrical figures presented in Figure 2 (left). The dependent variable was the relative frequency with which the participants selected the different landmark-based instructions at the intersections.

Procedure

The online experiment was implemented and administered with the software package Limesurvey 1.85. Participants were first informed that the experiment was concerned with route directions in human wayfinding. Then they saw the survey map of the environment including the complete route and all landmarks. They received the following instruction: “Imagine you must give verbal instructions to a person who is unfamiliar with this route, but needs to find his or her way to the target location. You will see different instructions for each intersection. Your task is to select the one that you think is the best.” An example for a map is given in Figure 3 (start at lower left with intersection 1 and compare with the following text).

FIGURE 3

Figure 3. Example of the street maps and the procedure of Experiment 1. For each decision point, participants had to select one of four route instructions each referring to one of the four landmarks at the intersection.

Participants then saw the map with the first intersection, the direction of the turn, and the four landmarks. Below the map, four instructions were presented, e.g., for the first intersection:

Turn right before the red triangle.

Turn right behind the blue circle.

Turn right behind the green hexagon.

Turn right before the yellow square.

The order of the four alternatives was randomized. Participants had to answer the following question: “Which of these instructions appears to be the best to you?” They made their choice by clicking with the mouse on one of the instructions. Then they saw the second intersection (and the previous one) with the landmarks and the direction of the turn and again had to select one of four instructions by mouse click. This procedure was repeated 16 times, one for each decision point (intersection). After finishing the 16 trials, the participants' demographic data were collected.

Results and discussion

Figure 4 shows how often participants selected each of the different landmark-based route instructions. Interestingly, in almost three quarters of the trials, participants selected instructions with landmarks located on the side of the turn and before the intersection (independent from color and shape). The second most frequent choice was again on the side of the turn, but behind the intersection. Route instructions with the landmarks on the opposite side of the turn were almost never selected. The statistical analysis using Kendall's W showed a significant agreement among participants in the rank order of instructions [W(3) = 0.571, p < 0.001].

FIGURE 4

Figure 4. Results of Experiment 1 given as the relative frequency [%] of chosen route instructions; the landmark positions are: “side of the turn, before the intersection” (bottom right); “side of the turn, behind the intersection” (top right); “opposite side of the turn, behind the intersection” (top left); “opposite side of the turn, before the intersection” (bottom left).

The results of Experiment 1 agree with our hypothesis that participants' landmark selection is not random. To the contrary, over the entire sample we found strong preferences for instructions with landmarks before the intersection in the direction of the turn. This finding agrees with some previous studies that demonstrate the particular importance of landmarks that are on the side of the turn (Waller and Lippa, 2007; Hölscher et al., 2011; Karimpur et al., 2016). We will return to this in the General discussion. A possible limitation of this exploratory study, however, was that people saw the environment from an allocentric bird's-eye perspective on a map. This, of course, is very different from the egocentric perspective we typically use when giving instructions to another person on a street. The two different perspectives are essential in almost all areas of spatial cognition research (e.g., Klatzky, 1998; Ekstrom and Isham, 2017). In the next experiment, we therefore presented the intersections and landmarks from an egocentric perspective.

Experiment 2: Egocentric presentation

In this experiment, the intersections were presented from an egocentric, i.e., observer-based perspective. To realize this, the same materials as in Experiment 1 were now presented in a video sequence which was generated from the SQUARELAND environment. Note that in an egocentric perspective the landmarks at the four corners of the intersection vary in their visibility as well as in their distance and orientation relative to the observer. Based on previous findings (e.g., Röser, 2017), our hypothesis was that participants should select the same landmarks as in Experiment 1 but that their preference for particular landmarks might be less pronounced than in the allocentric bird's-eye perspective of Experiment 1.