Multi-resolution Image Analysis for Vehicle Detection

Hilario, Cristina; Collado, Juan Manuel; Armingol, José Maria; de la Escalera, Arturo · 2005 · Crossref

DOI: 10.1007/11492429_70

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 computer vision-based Advanced Driver Assistance System (ADAS) for detecting vehicles, addressing the limitations of radar and laser sensors which possess narrow fields of view and struggle with lateral detection. While vision systems offer richer environmental descriptions, they face challenges in processing complexity and performance under adverse weather conditions. The authors adopt a top-down approach, utilizing a geometric model to locate vehicles, rather than bottom-up feature extraction or learning-based methods, to achieve robustness and real-time performance. The methodology defines a vehicle using a seven-parameter geometric model, including position, width, height, windshield and bumper positions, and roof angle. These parameters are constrained by a prior road detection module, which models road borders to restrict the search space and relate vehicle dimensions to road width. An energy function evaluates the fit of the model to the image by combining three factors: symmetry of vertical and horizontal edges, shape based on gradient and distance to edges, and the intensity of the vehicle’s shadow. The system employs a Gibbs distribution for the likelihood probability density function, framing detection as a Maximum A Posteriori estimation problem. To solve this optimization, a genetic algorithm is used to find the parameter set that minimizes the global energy function. To ensure real-time operation, the authors implement a multi-resolution approach using a Gaussian pyramid with resolutions of 160x120, 320x240, and 640x480 pixels. Information from lower-resolution levels guides the search at higher resolutions. This strategy reduces computational load significantly, requiring only 32 individuals in the genetic algorithm compared to 550 in previous work, reducing the average generation time from 24 ms to 0.16 ms on a Pentium 4 Mobile processor. The multi-resolution method also improves accuracy by mitigating errors caused by internal vehicle edges or cluttered environments that might deceive single-resolution algorithms. Experimental results on real images demonstrate successful detection of vehicles in front of the camera, as well as lateral vehicles and trucks. The system effectively handles various scenarios, though errors occur when rectangular environmental objects like buildings are mistaken for vehicle parts, or when vehicles are extremely close to the camera, resulting in under-detection. The study concludes that the proposed geometric model, combined with a multi-resolution genetic algorithm, provides a viable solution for real-time vehicle detection in intelligent vehicles, offering a complementary alternative to range-based sensors.

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.

• visual search