Adaptive two-phase thermal pedestrian detection using dual-tree complex wavelets and ga-optimized neural networks

Pedestrian detection in thermal imagery remains a difficult task due to inherently low image resolution, weather-dependent noise, and background regions that often share similar thermal signatures with human targets. Many existing handcrafted or feature-based approaches lack the adaptability needed to handle changing environmental conditions, leading to notable drops in accuracy under adverse weather. To overcome these limitations, this study presents an adaptive two-phase pedestrian-detection framework that dynamically modifies its detection pipeline based on the detected weather conditions of each scene. In the first phase, environmental categories are determined using features extracted from the Dual-Tree Complex Wavelet Transform (DT-DDCWT) along with a lightweight feedforward neural network. In the second phase, a weather-aware pedestrian-detection module is activated, which incorporates adaptive Regions of Interest (ROI) extraction, Genetic-Algorithm-based feature optimization, and  Genetic Algorithm (GA)-optimized Artificial Neural Networks (ANN) trained individually for each weather category.  Experiments were conducted on the OSU Thermal Pedestrian Dataset, which includes ۲۸۴ annotated thermal images captured under various weather conditions. A ۱۰-fold cross-validation protocol was employed to ensure unbiased evaluation and prevent data leakage, where each subset served once as the test partition while being excluded from training in its corresponding fold. The proposed framework achieved high and consistent performance across all folds, demonstrating its robustness under diverse environmental conditions.

کلیدواژه ها:

Pedestrian Detection ، Wavelet ، Neural Network ، Genetic Algorithm