Scholarlink

<p>Early gesture recognition aims to predict a gesture label from partially observed sequences. Traditional methods treat this as a sequential classification task, failing to leverage diverse relationships among negative categories, leading to potential misclassifications. To address this, we introduce Adaptive Category Exclusion with Proximal Policy Optimization (ACE-PPO), an RL-based framework that dynamically eliminates irrelevant categories to refine classification decisions. Integrating this into our proposed Reinforcement Learning-based Adaptive Gesture Recognition and Verification (RL-AGRAV) framework, we optimize gesture recognition accuracy using policy-based reinforcement learning. The exclusion mechanism is modeled as a mask operation applied to the classification probability output, allowing the RL agent to iteratively refine category selection through binary masks. Evaluated on the ASL Hand Sign and 20BN-jester datasets, our method achieves peak accuracies of 99.2% and 98.2% under complex conditions, improving early-stage recognition by 2.98% and 3.16% over baseline models, respectively. Parameter sensitivity analysis shows high AUC values of 88.14% and 87.53% with exclusion. Experimental results confirm that RL-AGRAV outperforms state-of-the-art methods, including MS-RNN, DeepSCN, MemLSTM, and MTS-SVM. ACE-PPO enhances early recognition by minimizing category ambiguity, while RL-AGRAV ensures reliable classification through confidence-based verification.</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

Dynamic hand gesture recognition via reinforcement learning with adaptive category exclusion

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Vikas Bhatt,
  • Ratnakar Dash

摘要

Early gesture recognition aims to predict a gesture label from partially observed sequences. Traditional methods treat this as a sequential classification task, failing to leverage diverse relationships among negative categories, leading to potential misclassifications. To address this, we introduce Adaptive Category Exclusion with Proximal Policy Optimization (ACE-PPO), an RL-based framework that dynamically eliminates irrelevant categories to refine classification decisions. Integrating this into our proposed Reinforcement Learning-based Adaptive Gesture Recognition and Verification (RL-AGRAV) framework, we optimize gesture recognition accuracy using policy-based reinforcement learning. The exclusion mechanism is modeled as a mask operation applied to the classification probability output, allowing the RL agent to iteratively refine category selection through binary masks. Evaluated on the ASL Hand Sign and 20BN-jester datasets, our method achieves peak accuracies of 99.2% and 98.2% under complex conditions, improving early-stage recognition by 2.98% and 3.16% over baseline models, respectively. Parameter sensitivity analysis shows high AUC values of 88.14% and 87.53% with exclusion. Experimental results confirm that RL-AGRAV outperforms state-of-the-art methods, including MS-RNN, DeepSCN, MemLSTM, and MTS-SVM. ACE-PPO enhances early recognition by minimizing category ambiguity, while RL-AGRAV ensures reliable classification through confidence-based verification.