Discrete features for rapid pedestrian detection in infrared images

Olmeda, Daniel; Armingol, Jose Maria; de la Escalera, Arturo · 2012 · Crossref

DOI: 10.1109/iros.2012.6385928

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 challenge of rapid pedestrian detection in far-infrared images for advanced driver assistance systems, particularly in low-visibility conditions. Traditional methods often rely on temperature thresholding, which requires calibrated sensors, or dense sliding-window approaches using features like Histogram of Oriented Gradients (HOG), which are computationally expensive and sensitive to the wide temperature variations inherent in infrared imagery. To overcome these limitations, the authors propose a method based on discrete phase congruency features. This approach aims to identify unique keypoints invariant to illumination and scale, allowing the system to focus computational resources only on regions of interest (ROIs) likely to contain pedestrians, thereby significantly reducing processing time. The methodology consists of two primary stages. First, the system extracts discrete keypoints by analyzing phase congruency, a measure of local symmetry in the frequency domain that is independent of image contrast. Using Gabor filters in the Fourier domain, the algorithm identifies corners and generates descriptors based on histograms of phase congruency orientations. These descriptors are matched against predefined body-part models to generate candidate ROIs. Second, these ROIs are classified using a Support Vector Machine (SVM). The feature vector for classification is formed by concatenating histograms from non-overlapping blocks within the ROI. The system was trained on a dataset of 5,000 manually labeled samples, covering various temperatures and distances, with parameters optimized for kernel type and margin softness. Experimental results demonstrate that the proposed method achieves high detection accuracy while maintaining low computational cost. The SVM classifier performed best with a Radial Basis Function kernel, though a quadratic kernel offered nearly comparable performance with faster computation. The system achieved an average processing time of 60 milliseconds per frame (16 fps) on an Intel Core 2 Duo processor, enabling real-time operation. When compared to standard HOG algorithms trained on the same database, the proposed approach showed superior performance in detection error trade-off curves. Additionally, testing on the OSU Thermal Imagery Database yielded nearly perfect results. The system successfully tracked pedestrians in live urban driving scenarios, handling challenges such as partial occlusion and motion-induced ghosting, although false positives increased in images captured at very high temperatures. The significance of this work lies in its ability to provide robust, real-time pedestrian detection using low-resolution, uncalibrated, and non-refrigerated microbolometer sensors. By leveraging phase congruency, the method achieves invariance to the sensor’s temperature sensitivity and wide infrared spectrum variations, which typically degrade gradient-based features. This makes the system highly suitable for night-time driving safety applications where computational power is limited and environmental conditions vary widely. The authors conclude that this discrete feature approach offers a viable alternative to dense scanning methods, balancing detection accuracy with the speed required for embedded automotive systems.

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.

• vru conspicuity