Robust Human Pose Estimation and Action Recognition Over Multi-level Perceptron

This research proposes a novel method that combines human pose estimation (HPE) and sustainable event classification (SEC), focusing on skeleton and context-aware feature extraction for precise activity recognition through machine learning techniques. In contrast to traditional state-of-the-art approaches, our method improves posture and event identification in daily activities by integrating individual detection, skeletal structure creation, and advanced feature extraction methods. The process begins with video-to-image conversion, followed by sliding window application for gradient-magnitude images and background removal during preprocessing. Human detection is achieved using Otsu's thresholding, and key points are identified with the Shi-Tomasi algorithm. The pose estimation step generates human skeletonization, and feature extraction is performed using color histograms, Harris corner detection, and Speeded-Up Robust Features (SURF), optimized through Hill Climbing. Classification using a multi-level perceptron on the UCF-101 and Sports Videos in the Wild (SVW) datasets achieves accuracies of 86.9% and 86.7%, respectively. Our method's unique combination of advanced techniques represents a significant enhancement over existing solutions, with promising applications in areas such as augmented reality, service robotics, e-health, fitness, and security surveillance.