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

Human posture estimation and recognition are crucial for analyzing human activity and enhancing life logging. This research introduces a novel approach that integrates human pose estimation (HPE) and sustainable event classification (SEC) with advanced skeleton and context-aware feature extraction, leveraging machine learning techniques for precise activity recognition. Unlike traditional SOTA methods, our approach enhances posture and activity event identification in daily life by seamlessly combining individual detection, skeletal framework establishment, and sophisticated feature extraction techniques. Our workflow includes converting video footage to images, applying sliding window techniques for gradient-magnitude images, and performing background removal in pre-processing. Human detection is achieved using Otsu's thresholding, while key points are detected with the Shi-Tomasi algorithm. The pose estimation process achieves human skeletonization, with features extracted using color histograms, Harris corner detection, and Speeded-Up Robust Features (SURF), followed by optimization with Hill Climbing. Classification, performed using a multi-level perceptron on UCF-101 and Sports Videos in the Wild (SVW) datasets, yields accuracies of 86.7% and 86.4%, respectively. This method's unique integration of advanced techniques positions it as a significant advancement over existing SOTA approaches, with promising applications in augmented reality, service bots, e-health fitness, and security surveillance.