Improving roundabout modelling using drone video analytics
This project proposes the development of evidence-based parameter estimation methods to improve Main Roads Western Australia’s roundabout modelling practice and operational guidelines by accounting for various local conditions such as geometry, topography, location type (residential, industrial, rural etc.), traffic mix, and seasonality, as well as driving behaviour. The data will be used to develop dedicated roundabout models for Aimsun at micro-, meso- and macroscopic levels.
Models play a vital role in supporting decision-making at both strategic and operational levels in the transport industry. In this project, we focus on roundabouts, where significant delays on arterial roads occur. Designers rely on traffic models to test design performance, so the quality of model predications directly affects the quality of roundabout design. Data is the foundation of modelling but conventional manual traffic surveys are deficient in both quality and quantity.
Although a wide range of sophisticated software tools for traffic modelling have been developed over the years, the lack of abundant high-quality data hinders model calibration, validation, and continuous development to account for changing driving behaviour and local conditions.
This project addresses both quality and quantity problems in traffic data by applying the latest drone video analytics technology developed by University of Western Australia (UWA) researchers to inform and improve roundabout modelling.
Project objectives
Existing traffic modelling software packages provide quick and cost-effective ways to model roundabouts. All models need to reflect specific site conditions through appropriate inputs and parameters that are based on locally collected data.
Traditionally, traffic surveys rely on manual counting, which is expensive, inconsistent, unreproducible and error prone. Poor data jeopardise model accuracy and therefore design quality, resulting in over- or under-engineered roundabouts, wasting taxpayer dollars, exacerbating congestion and/or reducing road safety.
Even when videos are used, the recordings are manually processed, with no possibility to guarantee accurate/complete data. Moreover, certain parameters of the models cannot be directly measured (e.g., critical gap). Statistical methods have been developed (e.g. Wu, 2012; Troutbeck, 2014), but they often require large amounts of data, previously not available.
Roundabouts are particularly difficult due to their geometry and the complex interactions between different movements. This is exacerbated by extensive anecdotal evidence that drivers behave inconsistently at roundabouts, aspects linked to attention and different risk attitudes and driving skills.
Current methods assume drivers accept the shortest safe gap, despite conflicting evidence. Previous work (Wu, 2012; Troutbeck, 2014) is also inconclusive on the relationship between the distribution assumptions and variance of the critical gap. Troutbeck (2014) offered evidence from simulated data that the maximum likelihood method accurately predicted the mean critical gap and its variance, whereas Wu’s (2012) approach underestimated the variance (p.83).
Conventional manual surveys cannot track individual vehicles moving across intersections especially roundabouts due to the complex movement patterns, sheer volume of vehicles and limited line of sight. Therefore, they can only observe each approach separately and consequently often miss important information on interactions.
Gap acceptance parameters are affected by geometric features, circulating flows, speeds of the entry, weather and driver behaviour; yet they are neither explicitly modelled nor included in modelling guidelines because of the lack of detailed data. This sometimes results in unexpected traffic behaviours (priority sharing or even a reversal of priority) – see Main Roads (2021) Operational Modelling Guidelines, 2021: p.45. Also, due to equipment limitations, precision of time measurements has been minimal, with rounding to the nearest full second being used as a compromise.
The deleterious effects of these approximations were documented by Main Roads (2021) Operational Modelling Guidelines, which recommends higher accuracy to avoid large discrepancies in saturation flow estimation. Due to the limitations of manual surveys and the absence of explicit guidelines on choosing parameters modellers regularly have to heuristically adjust inputs or fall back to software default parameter values, which are often based on expert opinions or historical data derived elsewhere that are non-representative of local conditions.
The problem becomes more pronounced when modelling new (greenfield) sites, with no observational data to calibrate the model. The consequence is either an over-engineered intersection, which is wasteful; or an intersection with insufficient capacity, that is expensive to fix.
Similar problems also appear during road construction, when road authorities are required to minimise traffic disruptions through the provision of ‘good enough’ temporary intersections. Model accuracy is even more challenging because people seem to behave more cautiously at temporary sites due to unfamiliarity (Walker & Calvert, 2015).
The end results are chaos, driver frustration and possibly unsafe situations. We propose to develop evidence-based parameter estimation methods that can improve model results for roundabouts by accounting for various local conditions such as geometry, topography, location type (residential, industrial, rural etc.), traffic mix, and seasonality, as well as driver behaviour.
This project will address the problem by using the large amount of vehicle trajectory data we will collect over about 50 roundabouts to extract data as the training set and about 10 roundabouts to extract data for the test set. The video analytics (VA) technology developed by UWA researchers will be used to extract the traffic data from videos. VA is a subset of computer vision that applies artificial intelligence (AI) to automatically extract spatiotemporal data from videos.
It has only recently become practical because of the latest developments in multiple technologies including drones, computer vision, machine learning, and computer hardware (GPU – Graphics Processing Unit). When applied to traffic, each vehicle’s movement is automatically tracked to generate its trajectory, which can be used to analyse detailed vehicle interactions or aggregated traffic flow statistics. To the best of our knowledge, a dataset of similar nature is not openly available.
Although there are other commercial VA service providers, most of them can only generate simple data (e.g. vehicle counts and speed) which is not sufficient for our in-depth modelling analysis. Our software extracts data according to traffic engineering standards used in Australia, rendering data readily available for subsequent traffic analysis and modelling.
The UWA VA pipeline generates both disaggregated trajectory data of each vehicle in the scene and aggregated data such as vehicle counts and classification, critical gap analysis, delay time, queueing parameters, speed, passenger car unit (PCU) value etc. (Figure 1), which can be enhanced by secondary data such as geometry, seasonality, location type etc. Development of dedicated roundabout models for Aimsun Next to achieve better representation of driver behaviour at roundabouts and improve its model accuracy at micro-, meso- and macro- levels.
In this project, an evidence-based approach will allow Main Roads’ improvement of roundabout modelling practice and guidelines. Main Roads currently uses SIDRA modelling software which is one of the most popular intersection modelling packages for such a purpose, but its default parameters were based on data gathered elsewhere and often do not work well for WA conditions.
Extensive real traffic data gathered locally, from a wide variety of settings, combined with the novel analytic techniques will overcome the limitations of current practice: inconsistent results, limited behavioural content and ad-hoc adjustments. The results will help Main Roads update its Operational Modelling Guidelines and provide clearer instructions to modellers in terms of what parameters it should use under different conditions.
References
Main Roads, 2021, Operation Modelling Guidelines v.2.0
Troutbeck, RJ, 2014, Estimating the Mean Critical Gap, Transportation Research Record 2461, 76-84.
Walker, G, & Calvert, M, 2015, Driver behaviour at roadworks, Applied Ergonomics, 51, 18-29.
Wu, N, 2012, Equilibrium of Probabilities for Estimating Distribution Function of Critical Gaps at Unsignalized Intersections, Transportation Research Record, 2886, 45-55.
Please note …
This page will be a living record of this project. As it matures, hits milestones, etc., we’ll continue to add information, links, images, interviews and more. Watch this space!
Discover more from iMOVE Australia Cooperative Research Centre | Transport R&D
Subscribe to get the latest posts sent to your email.
