Novel Trajectory Prediction Algorithm Using a Full Dataset: Comparison and Ablation Studies

Abstract

Autonomous vehicles (AVs) have been lauded as the next stage in transportation for decades, although they suffer from several shortcomings, including interactions with pedestrians (Tabone et al. 2021). Previous studies have focused on investigating communication techniques between autonomous vehicles and pedestrians or on the interpretation of pedestrian actions (de Clercq et al. 2019; Stanciu et al. 2018; Habibovic et al. 2018). A recent informative review suggested several priorities for research on AV-pedestrian interactions, including the requirement for universal algorithms and large-scale datasets, covering a range of road types, car designs and pedestrian demographics (Rasouli and Tsotsos 2020). This study contributes to these requirements, by investigating perceived safety and road crossing decisions in response to traditional and autonomous vehicles, at a range of crossing points. Investigating the pedestrian-AV interactions can be difficult, given the relative expense and danger of physical experiments including an unpredictable machine and vulnerable participants. This has led to previous studies on pedestrians and AVs performing stated-preference (SP) approaches, rather than physical experiments, to investigate relevant details surrounding situational outcomes and AV design. There is a wealth of knowledge surrounding the design of SP surveys, however there are questions remaining surrounding the best medium in which to present them, and how much of the relevant information is interpreted by the participant. Visualization techniques such as images and videos are often used as a dense, easily interpretable medium to provide participants with exact replicas of a situation. However, such data is usually presented in 2-dimensional formats (e.g. a computer), which necessarily reduces some of the information available, and can also reduce the level of immersion experienced by the participant. Virtual reality (VR) is an established research tool that provides an immersive environment while allowing user interaction. However, there remain numerous questions surrounding the validity of using VR and the extent of its benefits as a data generating tool (Arellana et al. 2020; Mokas et al. 2021; Rossetti and Hurtubia 2020). As a result of these limitations, it is still unknown how much the level of immersion can influence a participant’s SP responses. This study investigates participant SP responses to scenarios surrounding autonomous vehicles and road-crossing behaviour using both computer screens and VR headsets as experiment paradigm. This will investigate the relationship between perceived safety and road crossing choices as a function of car type (autonomous or traditional vehicles) and of nearby infrastructure (pedestrian crossings, signalised crossing, and basic road crossing). This builds on the previous studies that have performed SP investigations into pedestrian-AV interactions using VR (Farooq, Cherchi, and Sobhani 2018; Nuñez Velasco et al. 2019) but extend it by performing identical experiments on computer screens in addition to within VR headsets. Thus, we will detect any differences in the level of participant immersion, and crucially, any differences in participant responses. In this study, participants were shown a series of 360° videos, filmed from the perspective of a pedestrian walking towards a crossing point on the pavement in London, UK. Within these video clips, there are either traditional cars, or AVs. The AVs are easily identifiable with numerous sensors (e.g. LIDAR) on the roof of the vehicle. The videos stop just as the pedestrian has the option to cross or wait for the vehicle to drive past. The videos show scenarios for zebra crossings, signalised crossings, and simple road crossings, with each scenario being repeated (in identical videos, as far as practicable) for traditional human-driven vehicles and AVs. The participant was then asked to respond to several SP questions, detailing their impressions and their hypothetical responses. Finally, after all six videos had been viewed, the participants were asked to rate their level of immersion, using the IPQ immersion questionnaire. The study is ongoing, but at the moment of writing there have been 150 participants (we will continue to recruit up to 350) for computer screens, 100 (200) participants for VR headsets, of which 80 (180) participants took part in both. These participants were sampled from members of the public and from staff/student populations. The members of the public were recruited using a social media advertising campaign, as well as in-person engagement events run at major institutions around the United Kingdom. The staff/student population were recruited over a month-long period, where they were invited to perform both the VR and the desktop survey. This study will analyse the difference between SP surveys performed on computer screens and within VR environments, using simple difference testing based on the levels of immersion (as measured by the IPQ score), and on the responses. It will then continue to assess any differences in perceived safety based on crossing type, vehicle type and experimental paradigm. Finally, this study produces structural equations relating the choice to cross a road with the perceived safety, participant demographic, crossing type, vehicle type, and further predictors. Provisional data shows being more immersed in VR does not translate into a difference in perceived safety but distinguishes between the respondents’ interactions with autonomous vehicles and traditional vehicles.

Publication
2022 7th International Choice Modelling Conference (ICMC 2022)