开放场景睡姿识别方法：基于投票的新颖性检测

Open-Set Scenarios Sleeping Posture Recognition Method: Voting-Based Novelty Detection

摘要: 传统的睡眠姿势识别模型通常是在已知明确类别的数据集上训练的，而这种基于封闭集识别的方法在遇到一些奇特罕见的睡姿时容易将其误分类为已知的睡眠姿势，从而导致睡姿识别系统在开放环境下的可靠性与适用性大大降低。为了解决这个问题，提出一种基于开放集识别的睡姿检测模型。该模型主要完成两个任务：标准监督分类和新颖性检测。当样本被输入到模型时，分类器可以根据样本的新颖性得分，采用阈值比较，判断输入样本是未知或者识别为已知类别。模型分别在压力图和RGB图上做了验证，实验结果表明在开放场景下，该模型不仅能够准确识别已知的睡姿类别，识别率达到99%，而且能够有效地甄别出未知的睡姿，AUROC值达到92%，AUPR值达到95%。

Abstract: Traditional sleep posture recognition models are usually trained on datasets with known and well-defined categories. However, this closed-set recognition method tends to misclassify some unusual and rare sleep postures as known sleep postures, which greatly reduces the reliability and applicability of the sleep posture recognition system in open environments. To address this issue, a novel open-set sleep posture detection model is proposed. This model performs two tasks: standard supervised classification and novelty detection. When a sample is input into the model, the classifier can determine whether the input sample is unknown or recognized as a known category based on the sample’s novelty score, using a threshold comparison. The model is validated on both pressure maps and RGB images, and experimental results show that in open scenarios, the model not only accurately recognizes known sleep posture categories with a recognition rate of 99%, but also effectively identifies unknown sleep postures, with an AUROC of 92% and an AUPR of 95%.