Scholarlink

<p>Smart cities must have proper parking management to minimize traffic, air pollution, and delays by commuters. The traditional sensor-based systems are limited in terms of scalability, cost, and maintenance, which highlights the necessity of smart vision-based systems. This paper introduces a Hybrid Federated Smart Parking Framework (HFSPF) that integrates YOLOv10-based spatial slot localization with Temporal Convolutional Networks (TCN) and a Transformer encoder to predict short-term occupancy. Whale Optimization Algorithm (WOA) is an adaptive control method that adjusts the major hyperparameters of the spatial and temporal modules in order to increase the convergence and accuracy. Privacy protection is attained via the Adaptive Attention-Guided Differential Privacy (AAG-DP) mechanism that employs the Rényi Differential Privacy (RDP), which will selectively mask sensitive gradients with utility preservation. Moreover, Bayesian Federated Aggregation (BFA) is resistant to client heterogeneity and adversarial or noisy updates. The performance of the framework is shown to be better with experimental performance on benchmark datasets with 96.8% mAP and 97.1% accuracy on PKLot and 97.2% mAP and 97.4% accuracy on CNRPark-EXT in detection tasks. In the case of temporal forecasting, the hybrid TCNTransformer model recorded an RMSE of 0.141 and MAE of 0.118 on the SFpark dataset, which was better than the current baselines. The framework is efficient on edge devices, and it can provide real-time, privacy-conserving parking control on resource-constrained smart city resources.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Secure and federated vision-based parking management using a hybrid deep learning and privacy-preserving framework for smart cities

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Jamal Alotaibi

摘要

Smart cities must have proper parking management to minimize traffic, air pollution, and delays by commuters. The traditional sensor-based systems are limited in terms of scalability, cost, and maintenance, which highlights the necessity of smart vision-based systems. This paper introduces a Hybrid Federated Smart Parking Framework (HFSPF) that integrates YOLOv10-based spatial slot localization with Temporal Convolutional Networks (TCN) and a Transformer encoder to predict short-term occupancy. Whale Optimization Algorithm (WOA) is an adaptive control method that adjusts the major hyperparameters of the spatial and temporal modules in order to increase the convergence and accuracy. Privacy protection is attained via the Adaptive Attention-Guided Differential Privacy (AAG-DP) mechanism that employs the Rényi Differential Privacy (RDP), which will selectively mask sensitive gradients with utility preservation. Moreover, Bayesian Federated Aggregation (BFA) is resistant to client heterogeneity and adversarial or noisy updates. The performance of the framework is shown to be better with experimental performance on benchmark datasets with 96.8% mAP and 97.1% accuracy on PKLot and 97.2% mAP and 97.4% accuracy on CNRPark-EXT in detection tasks. In the case of temporal forecasting, the hybrid TCNTransformer model recorded an RMSE of 0.141 and MAE of 0.118 on the SFpark dataset, which was better than the current baselines. The framework is efficient on edge devices, and it can provide real-time, privacy-conserving parking control on resource-constrained smart city resources.