Novel Method for Vehicle and Pedestrian Detection Based on Information Fusion

García, Fernando; de la Escalera, Arturo; Armingol, José María · 2013 · Crossref

DOI: 10.1007/978-3-319-02332-8_8

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 novel high-level data fusion method for detecting and classifying vehicles and pedestrians in road environments, aimed at improving Advanced Driver Assistance Systems (ADAS). Motivated by the high fatality rates in traffic accidents and the limitations of single-sensor systems, the authors propose a system that combines information from a 2D laser scanner and a computer camera. This approach seeks to overcome the specific weaknesses of each sensor—such as the high false positive rate of laser scanners and the unstructured information limitations of vision systems—to provide a reliable detection system suitable for real-world road safety applications. The methodology employs independent low-level classifiers for each sensor, followed by a high-level fusion stage. The laser scanner, mounted on the vehicle bumper, compensates for vehicle movement using GPS and inertial measurement data. Obstacle shapes are reconstructed using polylines, and classification is performed via pattern matching. Vehicle detection relies on the delay pattern of laser spots caused by vehicle movement, while pedestrian detection uses a pattern based on the angles between polylines representing leg positions. The computer vision system utilizes regions of interest provided by the laser scanner to reduce computational cost and false positives. Vehicle detection uses Haar-like features with cascade classifiers, while pedestrian detection employs Histogram of Oriented Gradients (HOG) features. The fusion stage uses a Kalman Filter to estimate obstacle movement and Global Nearest Neighbors (GNN) to associate detections with existing tracks. An M/N policy manages track creation and elimination, treating non-consolidated tracks (detected by only one sensor) as potential false positives. The system was tested in urban and interurban scenarios using more than 10,000 frames of real road data. Results demonstrated that the fusion approach significantly enhanced performance compared to individual sensor outputs. While the laser scanner alone achieved a high positive detection rate, particularly for vehicles, it suffered from a high number of misdetections. Conversely, the vision system had a lower false positive rate but struggled with detection consistency. The fused system successfully combined the strengths of both, reducing false positives through the reliability of the laser data and improving detection coverage through the visual data. The authors note that the vision system was trained specifically to minimize false positives given the laser scanner's tendencies, and that even in worst-case scenarios, vehicles were detected within one or two frames. The study concludes that high-level data fusion effectively overcomes the inherent limitations of individual sensors, providing a robust detection system ready for integration into real road applications. Although the cost of laser scanners remains a barrier for widespread adoption, the technology is already present in modern vehicles for obstacle avoidance. The work demonstrates that combining independent classifiers allows for a system that is both reliable and capable of functioning even if one sensor becomes unavailable, thereby meeting the stringent requirements of road safety 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.

• distraction detection algorithms