Author Affiliations
School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, Jiangsu 221116, Chinashow less
Fig. 1. Method for identifying dangerous actions of personnel in petrochemical scenes based on machine vision
Fig. 2. Process of skeleton-based human action recognition
Fig. 3. Skeleton-based human action recognition network
Fig. 4. Judgment method of the spatial position relationship between human and cell phone
Fig. 5. Judgment result of the spatial position relationship between human and cell phone
Fig. 6. Example of cell phone call action data set
Fig. 7. Loss value and accuracy rate
Fig. 8. Test result
Fig. 9. Confusion matrix (only skeleton information)
Fig. 10. Confusion matrix (adding object detection)
Fig. 11. Experimental results of simulated petrochemical scene
Fig. 12. Experimental results of actual petrochemical scene
Fig. 13. Examples of smoking action data set
Fig. 14. Results of smoking action recognition
Action No. | Action label |
---|
0 | walking | 1 | walking_suspected_calling | 2 | walking_and_calling |
|
Table 1. Initial action categories
Action label | Recall /% | Precision /% |
---|
walking | 98.07 | 95.50 | walking_suspected_calling | 71.47 | 75.90 | walking_and_calling | 74.51 | 70.66 |
|
Table 2. Action recognition results (before adding the object detection module)
Action label | Recall /% | Precision /% |
---|
walking | 98.07 | 96.58 | walking_suspected_calling | 99.21 | 85.78 | walking_and_calling | 81.92 | 99.18 |
|
Table 3. Action recognition results (after adding the object detection module)