Research and Design of Human Behavior Recognition Method in Industrial Production Based on Depth Image

Abstract

Correct assembly behavior in industrial production is a direct means to ensure production quality and efficiency. Aiming at the problems of worker misoperation or lack of important assembly steps in the production and assembly process, a human behavior recognition method based on ToF camera is proposed. The method segments the extracted depth motion maps (DMMs) according to the differences between frames, and extracts histogram of oriented gradient (HOG) descriptors and multiscale grayscale count (MGC) descriptors as local features. Then a hierarchical DMMs multi-classifier recognition framework is built based on stacking strategy, combining support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF) and XGBoost classifiers, achieving 98.2% accuracy on MSR ACTION 3D dataset and 87.1% accuracy on self-built dataset, respectively.