Armed Activity Recognition using Pose-Based Features with Machine Learning and Deep Learning

Automatic detection of "armed" activities commonly focuses on the detection of weapons, with the limitations of variety in shapes, types and colours of weapons. This study aims to perform AI-based identification of armed and unarmed activities using postural landmark features of the human body, without requiring weapon visibility. We conducted a comparative analysis on the postures by MediaPipe and YOLOv8-pose frameworks and their impact on the performance of conventional machine learning and deep learning classifiers. Posture landmarks were extracted separately using both frameworks on a total number of 3,866 images (1,934 armed and 1,932 unarmed activities). Conventional machine learning algorithms and Long Short-Term Memory (LSTM) algorithms were trained without data augmentation on GridSearchCV and the early stopping mechanism. The experimental results showed that the YOLOv8-pose-based posture feature landmarks provide higher armed pose activity classification performance than the MediaPipe-based landmarks. The optimized SVM trained on the YOLOv8-pose feature set achieved the highest test accuracy rate of 94.2%, whereas the most successful deep learning model, the YOLOv8-pose-based Bi-LSTM, reached an accuracy rate of 93.9%. Consequently, the findings demonstrate that feature set responsibility can outweigh model complexity of sequence learning-based algorithms in "armed pose" detection for even a lightweight SVM algorithm.