Scholarlink

<p>Accurate person re-identification (Re-ID) across independent or non-overlapping camera views remains a significant challenge in intelligent surveillance systems due to illumination variations, pose changes, occlusion, and background clutter. This paper presents a novel hybrid deep-learning framework that integrates Custom Convolution Neural Networks (CCNNs) for discriminative feature extraction with Radial Basis Function Neural Networks (RBFNNs) for robust nonlinear classification. The proposed CCNN–RBFNN architecture leverages the hierarchical spatial feature learning capability of CNNs while exploiting the fast generalization and localized kernel adaptability of RBFNNs to enhance identity recovery under complex real-world scenarios. The model is trained and optimized using data augmentation and hyperparameter tuning strategies to improve convergence stability and minimize inter-class confusion. Evaluations conducted on four benchmark datasets - CUHK01, PRID2011, MARS, and iLIDS-VID - demonstrate that the proposed hybrid model consistently outperforms contemporary state-of-the-art Re-ID approaches. Specifically, the framework achieves re-identification accuracies of 97.20%, 95.16%, 94.87%, and 93.25%, respectively, marking improvements of up to 4.3% over recent deep metric-learning methods. Experimental analysis confirms that the fusion of CNN feature hierarchies with RBF kernel decision boundaries not only enhances inter-camera identity matching but also significantly improves robustness against occlusion and illumination variation. The proposed approach therefore contributes a computationally efficient and scalable deep-learning paradigm for next-generation multi-camera surveillance and intelligent monitoring systems.</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

Vision based approach - radial basis kernels and deep-networks to recover person identity from independent camera views

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Deepak N. A.

摘要

Accurate person re-identification (Re-ID) across independent or non-overlapping camera views remains a significant challenge in intelligent surveillance systems due to illumination variations, pose changes, occlusion, and background clutter. This paper presents a novel hybrid deep-learning framework that integrates Custom Convolution Neural Networks (CCNNs) for discriminative feature extraction with Radial Basis Function Neural Networks (RBFNNs) for robust nonlinear classification. The proposed CCNN–RBFNN architecture leverages the hierarchical spatial feature learning capability of CNNs while exploiting the fast generalization and localized kernel adaptability of RBFNNs to enhance identity recovery under complex real-world scenarios. The model is trained and optimized using data augmentation and hyperparameter tuning strategies to improve convergence stability and minimize inter-class confusion. Evaluations conducted on four benchmark datasets - CUHK01, PRID2011, MARS, and iLIDS-VID - demonstrate that the proposed hybrid model consistently outperforms contemporary state-of-the-art Re-ID approaches. Specifically, the framework achieves re-identification accuracies of 97.20%, 95.16%, 94.87%, and 93.25%, respectively, marking improvements of up to 4.3% over recent deep metric-learning methods. Experimental analysis confirms that the fusion of CNN feature hierarchies with RBF kernel decision boundaries not only enhances inter-camera identity matching but also significantly improves robustness against occlusion and illumination variation. The proposed approach therefore contributes a computationally efficient and scalable deep-learning paradigm for next-generation multi-camera surveillance and intelligent monitoring systems.