PE-USGC: Posture Estimation-Based Unsupervised Spatial Gaussian Clustering for Supervised Classification of Near-Duplicate Human Motion

Near-duplicate human motion classification presents significant challenges due to the subtle differences and high similarity between actions. This paper introduces a posture estimation-based Gaussian Mixture Model (GMM) clustering algorithm as an enhancement to traditional pixel-based Convolutional Neural Networks (CNNs). The CNN architectures evaluated include ResNet-18, SqueezeNet, DenseNet, and MobileNet, which are used to classify images based on pixel data of images extracted from videos of participants performing chopping, sawing, and slicing tasks. While these CNN models perform well in extracting deep hierarchical features from images, differentiating between near-duplicate tasks can be challenging due to a lack of context in cross-frame human motion. In contrast, the posture-based approach focuses on capturing the spatial and temporal patterns of human body landmarks during task execution, using posture landmark points to classify human motion. Notably, posture-based task classification outperformed pixel-based task classification by 7.2%, with a lesser demand for image frame rate to achieve better accuracy. As the frames per second (FPS) increased from 1 to 30, the accuracy of posture-based classification improved from 76.3% at 1 FPS to 96.97% at 30 FPS. Additionally, we evaluated our model using the UCF Sports Action dataset as a benchmark to compare with state-of-the-art human task classification methods. These comparative analyses highlight the strengths and limitations of each approach, demonstrating that integrating posture data can enhance the classification accuracy of near-duplicate human motion.