Scholarlink

<p>This study presents a comparative framework for human gait recognition, a critical application in security and surveillance engineering. The work focuses on evaluating and optimizing three classical feature descriptors—Scale Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), and Shi-Tomasi corner detection—combined with three classification models: Decision Tree, Random Forest, and Multi-Layer Perceptron (MLP). The AI contribution of this work lies in the hybrid integration of these handcrafted features, dimensionality reduction using Locality Preserving Projections (LPP), and rigorous benchmarking across multiple classifier-feature combinations. The proposed methodology is validated on the CASIA-A dataset using an 80:20 train-test split. Experimental results demonstrate that the Shi-Tomasi descriptor paired with the Random Forest classifier achieves the highest recognition accuracy of 82.13%. Evaluation metrics include accuracy, false positive rate (FPR), root mean squared error (RMSE), as well as additional statistical and runtime analyses. The findings underscore the viability of lightweight, interpretable machine learning pipelines for real-time gait recognition in surveillance environments.</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

Comparative Analysis of Feature Descriptors and Classifiers for Human Gait Recognition

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Ahmed Alkhayyat,
  • Junainah Abd Hamid,
  • B. Jayaprakash,
  • Krishan Kumar Sah,
  • Mandeep Kaur Chohan,
  • Amrita Singh,
  • S. Srinadh Raju,
  • Rajesh Singh

摘要

This study presents a comparative framework for human gait recognition, a critical application in security and surveillance engineering. The work focuses on evaluating and optimizing three classical feature descriptors—Scale Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), and Shi-Tomasi corner detection—combined with three classification models: Decision Tree, Random Forest, and Multi-Layer Perceptron (MLP). The AI contribution of this work lies in the hybrid integration of these handcrafted features, dimensionality reduction using Locality Preserving Projections (LPP), and rigorous benchmarking across multiple classifier-feature combinations. The proposed methodology is validated on the CASIA-A dataset using an 80:20 train-test split. Experimental results demonstrate that the Shi-Tomasi descriptor paired with the Random Forest classifier achieves the highest recognition accuracy of 82.13%. Evaluation metrics include accuracy, false positive rate (FPR), root mean squared error (RMSE), as well as additional statistical and runtime analyses. The findings underscore the viability of lightweight, interpretable machine learning pipelines for real-time gait recognition in surveillance environments.