Dataset Construction from Naturalistic Driving in Roundabouts

Cuenca, Laura García; Guindel, Carlos; Aliane, Nourdine; Armingol, José María; Andrés, Javier Fernández · 2020 · DOAJ

DOI: 10.3390/s20247151

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the lack of publicly available, high-quality datasets specifically designed for modeling driver behavior in complex traffic environments, such as roundabouts. While existing autonomous driving datasets like KITTI and nuScenes focus on general environment perception, they often lack the specific maneuver-level data required for driver characterization in these dynamic scenarios. The authors propose the construction of an open-access dataset derived from naturalistic driving, aiming to facilitate research into autonomous vehicle control and driver modeling in roundabouts. The methodology involves a custom data acquisition system installed in a vehicle, comprising a smartphone for GPS, acceleration, and video recording (30 fps), and an Arduino board for capturing driver inputs (pedals, blinkers, steering wheel rotation) at 2 Hz. Data were collected from multiple drivers navigating various roundabouts in Madrid, Spain. The dataset construction process segments each roundabout into three phases: 100 meters before entry, the internal maneuver, and 100 meters after exit. These phases are spatially sampled into smaller sections (20-meter segments for approach/exit and 45-degree angular segments for the interior). Feature engineering combines on-board sensor data with cartographic information from OpenStreetMaps (e.g., diameter, lane count) and processed video data. A Faster R-CNN deep learning model, fine-tuned on the nuScenes dataset, was used to detect objects and estimate their orientation in video frames. This pipeline generated four traffic-related parameters for each segment: crowdedness (object count), presence of vulnerable road users (VRUs), distance to the closest vehicle, and dominant orientation of surrounding vehicles. The resulting dataset comprises 33 routes containing 377 roundabouts, organized into a CSV file with attributes for each spatial segment. Statistical analysis reveals vehicle speeds ranging from 0 to 69 km/h (average 36.77 km/h) and roundabout diameters from 13 to 103 meters. The video processing pipeline demonstrated high accuracy, with mean absolute errors below 1 unit for crowdedness and orientation estimates, and a median absolute error of 1.74 meters for distance estimation. Precision for vehicle detection reached 89.9%, while VRU detection showed high precision (87.9%) but lower recall (66.3%). The dataset captures diverse traffic conditions, including varying levels of crowdedness and VRU presence, with most roundabouts featuring two lanes. The significance of this work lies in providing a specialized, pre-processed dataset that allows researchers to focus on high-level maneuver modeling without the computational burden of raw sensor processing. By making this data open-source, the authors aim to support the development of expert systems for autonomous driving that can replicate human-like behavior in complex intersections. This resource fills a gap in the field by offering specific, labeled data for roundabout maneuvers, which are critical for safe and efficient autonomous navigation.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• naturalistic crash near crash

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: dataset resource, tool software