SAFETY HELMET WEARING DETECTION BASED ON AN IMPROVED YOLOV3 SCHEME

During the construction process, ensuring construction safety is an important link to improve production efficiency, enhance corporate efficiency, and ensure employee safety. Real-time checking whether workers wear safety helmets is naturally a key task in ensuring safe production. In order to reduce the incidence of safety accidents caused by not wearing a helmet, a helmet wearing detection method based on the improved YOLO (You Only Look Once) v3 algorithm is proposed. The improvement of this paper is reflected in two aspects: one is the improvement of the YOLOv3 algorithm itself, and the other is the improvement of the overall process. Aiming at the improvement of YOLOv3, this paper improves the feature fusion steps of YOLOv3, where we use upsampling to fuse high-level features with low-level features. From the perspective of the overall process, this paper first calibrates the relevant data set, and divides the data set into four categories. For the calibrated data, transfer learning is used to train the improved YOLOv3 network. Then, these parameters and model are used to detect the categories and positions of human figures and helmets on surveillance video data. Finally, we calculate several intersection-over-unions among the four detected types, and use this to judge whether the worker is wearing a helmet correctly. Experimental results show that the proposed algorithm satisfies the real-time performance of the detection task in the helmet wearing detection, and has a higher detection accuracy rate. The mAP (mean average precision) reaches 95.13%, and this detection accuracy rate is higher than that of the traditional SSD (single shot multibox detector), Faster R-CNN (regional-convolutional neural network) and YOLOv3.