A Machine-Learning Classification Approach to Automatic Detection of Workers' Actions for Behavior-Based Safety Analysis
Seunghyun Han, Sang Hyun Lee, Feniosky A. Pena-Mora · 2012
About 80-90% of accidents are strongly associated with workers' unsafe behavior. Despite the importance of worker behavior measurement, it has not been applied actively in practice due to its time-consuming and painstaking associated tasks. For measuring and analyzing worker behavior, vision-based motion capture has recently been proposed as an emerging technique that requires no additional time or cost. In line with motion tracking that extracts 3D skeleton motion models from videos, this paper proposes motion classification techniques for automatic detection of workers' actions. The high-dimensional motion models (e.g., 78 variables in this study) that result from tracking are transformed into a 3D latent space to reduce the dimensions for recognition. To recognize motions in the 3D space, this paper applies supervised classification techniques for training the learning algorithms with training datasets where unsafe motions are labeled, and then classifying testing datasets based on the learning. As a case study, motions during ladder-climbing are tested. The results indicate that the proposed approach performs well at automatically recognizing particular motions in datasets. Thus the measured information has great potential to be used for enhancing safety by providing feedback and improving workers' behavior.