Context aided fusion procedure for road safety application

Garcia, Fernando; de la Escalera, Arturo; Armingol, Jose Maria; Jimenez, Felipe · 2012 · Crossref

DOI: 10.1109/mfi.2012.6343070

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 presents a context-aided multisensor fusion procedure designed to enhance vehicle detection and danger evaluation for road safety applications, specifically Advanced Driver Assistance Systems (ADAS). The research addresses the limitations of individual sensors: laser scanners provide robust obstacle detection but limited classification information, while computer vision offers rich data but suffers from lower reliability and higher false positive rates. By integrating these sensors with context information—such as vehicle physics, dynamics, and traffic safety metrics—the authors aim to create a reliable, multilevel fusion architecture that minimizes misdetections and accurately assesses collision risks. The experimental system utilizes three primary sensors: a SICK LMS-291-S05 laser scanner, a Point Grade CCD firewire camera, and an X-sens MTi-G GPS-inertial measurement unit. The fusion process operates across multiple levels. At the detection level, the laser scanner identifies candidate obstacles, which define Regions of Interest (ROIs) for the camera. This gating mechanism reduces computational load and filters visual false positives. Coordinate transformations align the laser and camera data with the vehicle’s reference frame. Vehicle classification is performed using Haar Cascade Classifiers on the visual ROIs. For tracking, an Unscented Kalman Filter manages data association and gating, employing a constant velocity model suitable for interurban roads. Track management logic distinguishes between "consolidated" tracks (confirmed by both sensors) and "non-consolidated" tracks, applying specific thresholds for track creation and deletion to further suppress false positives. Context information, including vehicle dimensions and movement physics, is used to filter impossible trajectories. At the situation assessment level, the system calculates braking distance and safety distance to evaluate danger. Braking distance is computed using worst-case friction coefficients and human reaction time assumptions, while safety distance is derived from the velocities of the host and target vehicles. Tests were conducted in interurban scenarios with low traffic, involving overtaking maneuvers. The results demonstrated that the standalone camera achieved 60.13% positive detections with 1.20% false positives, while the laser scanner achieved 84.09% positive detections with 5.30% false positives. The fused system improved overall performance to 90.15% positive detections with only 0.76% false positives. The study concludes that combining laser scanner trustability with computer vision classification, supported by context-aware fusion, yields detection rates comparable to widely used frequency radars but with significantly fewer false positives. The authors note that while the system performs well in interurban environments, future work must address the complexities of urban scenarios, where higher obstacle density and curved trajectories challenge current laser reconstruction and inertial movement prediction methods.

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