Enhanced obstacle detection based on Data Fusion for ADAS applications

Garcia, Fernando; de la Escalera, Arturo; Armingol, Jose Maria · 2013 · Crossref

DOI: 10.1109/itsc.2013.6728422

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 decentralized high-level data fusion architecture designed to enhance obstacle detection for Advanced Driver Assistance Systems (ADAS). The research addresses the need for highly reliable sensing technologies in road safety applications, where individual sensors often suffer from inherent limitations. Specifically, the study aims to combine data from laser scanners and computer vision systems to detect pedestrians and vehicles with greater trustability and accuracy than single-sensor approaches. The proposed system utilizes a decentralized fusion scheme where independent low-level detection subsystems process data from a laser scanner and a camera. The laser scanner identifies obstacles based on polyline reconstruction and movement patterns, classifying pedestrians via leg movement analysis and vehicles via specific motion signatures. The computer vision subsystem uses Haar-like features with cascade classifiers for vehicle detection and Histogram of Oriented Gradients (HOG) with Support Vector Machines (SVM) for pedestrian detection. These detections are fused at a high level using a Joint Probabilistic Data Association Filter (JPDAF) adapted for Multiple Target Tracking (MTT). This algorithm manages track creation, deletion, and updates, employing a constant velocity model and Kalman Filter for state estimation. The system was tested in real urban and interurban scenarios using over 10,000 frames. The results demonstrate that the JPDA-based fusion significantly outperforms both individual low-level detections and a Global Nearest Neighbor (GNN) fusion approach. For pedestrian detection, the JPDA system achieved a positive detection rate of 82.29% with a false positive error rate of 1.11%, compared to 77.69% and 3.11% for GNN, and lower performance for individual sensors. For vehicle detection, JPDA achieved a 92.03% positive rate with only 0.59% error, surpassing the GNN’s 88.25% positive rate and 2.59% error. The laser scanner alone showed high false positive rates due to limited information, while the camera struggled with missed detections. The JPDA algorithm effectively reduced false positives, particularly in challenging situations involving occlusions, merging groups, and crossings. The study concludes that high-level data fusion using JPDA provides a scalable and robust solution for ADAS applications. By combining the strengths of laser scanners and computer vision, the system overcomes the limitations of each sensor, such as the unstructured nature of visual data and the limited information from laser scans. The decentralized architecture allows for easy integration of future sensing devices, and the independence of subsystems ensures detection capability even if one sensor fails. The findings validate the effectiveness of JPDA in enhancing the reliability and accuracy of road user detection, meeting the stringent demands of safety-critical automotive applications.

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.

• adas effectiveness