Rivera Fernando Pujaico, Rodrigues Paulo Sergio, Fugita Oscar Eduardo Hidetoshi
Department of Electrical Engineering, University Center of FEI, SBC, SP, Brazil.
University Hospital - University of São Paulo, SP, Brazil.
Comput Biol Med. 2025 Jul;193:110350. doi: 10.1016/j.compbiomed.2025.110350. Epub 2025 May 21.
Body language is a crucial aspect of communication, as it helps observers interpret emotional states based on non-verbal cues. However, accurately classifying body language into distinct emotional categories remains a challenging task, particularly due to the limited availability of high-quality datasets focused on various body expressions, particularly in contexts such as the health sector. The computational problem addressed in this study arises from the scarcity of datasets suitable for training classifiers capable of distinguishing four categories of body expressions (negative, neutral, pain, and positive). To address this gap, we introduce the Body Emotion Recognition dataset, specifically designed to provide a robust foundation for training and evaluating body language classifiers. This dataset, created from images of individuals simulating the aforementioned categories, was utilized to test and analyze the performance of three distinct convolutional neural network approaches: one focused on facial images, another on body images, and a third on skeletal posture data represented as keypoints. Transfer learning from models pretrained on ImageNet and other datasets was employed to demonstrate the potential of the proposed dataset in achieving accurate classification of the four body expression categories. Our approach, tested on these models, has achieved varying accuracies depending on the type of data analyzed: a test accuracy of 96.25% when analyzing facial images, 95.59% when analyzing body images, and 67.53% when analyzing skeletal posture data represented as keypoints. These results highlight the quality of the dataset in carefully labeling the images and indicate the amount of information that can be extracted from these categories in the dataset, thus providing a valuable tool for future research in the area of health care.
肢体语言是沟通的一个关键方面，因为它有助于观察者根据非语言线索解读情绪状态。然而，将肢体语言准确分类到不同的情绪类别仍然是一项具有挑战性的任务，特别是由于专注于各种身体表达的高质量数据集有限，尤其是在医疗保健等领域。本研究中解决的计算问题源于缺乏适合训练能够区分四类身体表达（负面、中性、疼痛和正面）的分类器的数据集。为了填补这一空白，我们引入了身体情绪识别数据集，专门设计用于为训练和评估肢体语言分类器提供一个强大的基础。这个数据集由模拟上述类别的个人图像创建而成，用于测试和分析三种不同的卷积神经网络方法的性能：一种专注于面部图像，另一种专注于身体图像，第三种专注于表示为关键点的骨骼姿势数据。采用从在ImageNet和其他数据集上预训练的模型进行迁移学习，以证明所提出的数据集在实现对四类身体表达的准确分类方面的潜力。我们在这些模型上进行测试的方法，根据所分析的数据类型取得了不同的准确率：分析面部图像时测试准确率为96.25%，分析身体图像时为95.59%，分析表示为关键点的骨骼姿势数据时为67.53%。这些结果突出了该数据集在仔细标注图像方面的质量，并表明可以从数据集中的这些类别中提取的信息量，从而为医疗保健领域的未来研究提供了一个有价值的工具。
