Abstract: In order to address flaws of existing helmet-wearing detection model, such as its requirement of large sample data and inclination to false detection, a new detection model was proposed that combined human joint detection and Faster R-CNN. Then, OpenPose was utilized to locate positions of head and neck joints, and sub-image of small areas near helmet was extracted before it was detected with Faster R-CNN. Finally, spatial relationship between helmet and head/neck joints were analyzed to further verify whether it was worn correctly. The results show that this enhanced method can reduce error rate and improve its environmental adaptation effectively. And even with small sample data, its recall rate increases by more than 20% and detection accuracy by approximately 10%, significantly reducing demand on samples.

Key words: helmet-wearing detection, human joint, Faster R-CNN, sub-image, spatial relationship