Towards Fairer and More Accurate Real-Time Pedestrian Attribute Recognition for Enhanced Women’s Safety: A Domain-Adversarial Multi-Head Model with Agent-Based Reporting

Effective surveillance systems play a vital role in improving public security, most important among which are applications related to women’s security, where the capacity to correctly carry out Pedestrian Attribute Recognition (PAR) is of utmost importance. Current PAR models are susceptible to being thrown off by dataset bias, mainly gender bias due to training set imbalance, resulting in suboptimal generalization and incorrect prediction across groups. This paper solves the challenges above by creating a Domain-Adversarial Training for Multi-Head Pedestrian Attribute Recognition (DAMH-PAR) model. DAMH-PAR uses a domain-adversarial training method combined with a multi-head structure to provide specialized training to various sets of attributes. The model learns invariant domain features upon training independent "expert" heads per dataset, which are chosen during inference to produce detailed pedestrian descriptions. The usefulness of the model is established by its capacity to predict pedestrian attributes accurately from security feeds even under heavy lighting. Such details, combined with violence detection and proximity modules, forms critical contextual information for automated safety systems. Testing DAMH-PAR model on PETA and PA-100K datasets reveals significant performance improvement compared to existing top-performing benchmarks. The DAMH-PAR model achieves 90.50% Mean Accuracy on PETA and 94.31% accuracy on PA-100K, which is more than the previously established standards. The performance highlights the potential of the suggested methodology to develop more effective and unbiased PAR models for meaningful development in security and safety surveillance tasks.