Open Access

ARTICLE

Robust Pedestrian Detection in Surveillance Videos via Fractal-Geometric Regularization of Conditional Random Fields

Mohammadreza Nehzati^*

VMC MAR COM Inc. DBA Axiomera, Knoxville, TN, USA

* Corresponding Author: Mohammadreza Nehzati. Email:

Journal on Artificial Intelligence 2026, 8, 299-322. https://doi.org/10.32604/jai.2026.082982

Received 26 March 2026; Accepted 09 May 2026; Issue published 16 June 2026

Abstract

Pedestrian detection in surveillance environments remains fundamentally challenging due to three coupled phenomena: severe occlusion, extreme scale variation, and low-resolution imagery. While contemporary detectors achieve high frame rates on standard benchmarks, they exhibit systematic failures under partial visibility where geometric consistency becomes the primary discriminative signal. This paper introduces a hybrid probabilistic framework that integrates Conditional Random Fields (CRF) with fractal geometry regularization to enforce scale invariant shape priors in deep pedestrian detectors. The core mathematical insight is that human silhouettes exhibit self-similarity across scales—a property precisely characterized by fractal dimension, self-similarity coefficients, and non-dimensionality indices. We formulate these geometric features as differentiable regularizes that guide both network training and CRF inference. The complete pipeline comprises: (i) photometric normalization and denoising using adaptive Gaussian filtering, (ii) hierarchical feature extraction via CNN backbones, (iii) parallel fractal feature computation via GPU-accelerated box-counting algorithms, and (iv) dense CRF optimization with novel fractal-aware pairwise potentials. Extensive evaluation on three benchmarksCity Persons, KAIST infrared, and INRIA demonstrates consistent statistically significant improvements. Ablation studies isolate component contributions: CRF optimization provides +2.3% AP through spatial coherence, fractal regularization contributes +1.8% AP specifically under occlusion, and their combination yields synergistic +4.8% total improvement exceeding the sum of individual gains. This work method achieves 82.7% AP@0.5 on City Persons and 85.2% AP@0.5 on KAIST infrared at 5.6 fps on an RTX 3080 Ti. Statistical validation using bootstrapped confidence intervals (n = 1000) and paired t-tests with Bonferroni correction confirms significance (p < 0.001, Cohen’s d > 0.8). This work demonstrates that classical fractal geometry provides complementary geometric priors orthogonal to modern architectural advances, offering an interpretable, mathematically-grounded alternative for reliability-critical surveillance applications.

---

Keywords

---