Part based pedestrian detection based on Logic inference

Olmeda, Daniel; Armingol, Jose Maria; de la Escalera, Arturo · 2013 · Crossref

DOI: 10.1109/itsc.2013.6728421

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 detecting pedestrians in urban environments, particularly when they are largely occluded or only partially visible within the camera frame. The authors propose a method that fuses data from synchronized Visible Light (VL) and Far Infrared (FIR) cameras to improve detection robustness. The motivation stems from the complementary nature of these sensors: VL systems perform best under high illumination (daylight), while FIR systems excel in low-light or night conditions by detecting heat signatures. By combining these modalities, the system aims to provide reliable pedestrian detection for Intelligent Transportation Systems (ITS) and Advanced Driver Assistance Systems (ADAS) across varying environmental conditions. The methodology employs a constrained sliding window approach using Histogram of Oriented Gradients (HOG) for VL images and a contrast-invariant descriptor called HOPE for FIR images. Individual detections from both sensors are projected into a three-dimensional world coordinate system, assuming pedestrians stand on the pavement, and matched using the Munkres assignment algorithm. The core innovation is the use of a Markov Logic Network (MLN) to infer final detection confidence. The MLN combines binary detection results from VL and FIR sensors using logic rules, assigning weights to statements such as "if detected in VL, then pedestrian." To handle occlusion, the approach extends to part-based detection by dividing the detection window into subparts and training Support Vector Machine (SVM) classifiers for each cell. The MLN is further enhanced by automatically learning latent structures and rule weights from the data, allowing it to infer pedestrian presence from incomplete samples. Experimental results demonstrate that the logic combination of VL and FIR detections outperforms individual sensor performance, as shown in Detection Error Trade-off (DET) curves. The part-based model significantly improves classification accuracy for heavily occluded pedestrians compared to full-body rigid models. While the part-based approach yields slightly lower performance on fully visible pedestrians, it maintains high detection rates for samples with up to 50% lateral occlusion. The learned latent structures in the MLN allow the system to identify pedestrians based on visible body parts, effectively overcoming the limitations of traditional methods that require complete object visibility. The significance of this work lies in its ability to enhance pedestrian detection reliability in complex urban scenarios where occlusion and varying lighting conditions are common. By leveraging the complementary strengths of VL and FIR sensors through probabilistic logic inference, the proposed system offers a robust solution for safety-critical applications like autonomous driving and traffic surveillance. The findings suggest that integrating multi-spectral data with part-based modeling and logic networks can substantially improve the detection of vulnerable road users, potentially contributing to accident prevention and improved traffic management.

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