A helmet detection method with lightweight backbone based on yolov3 network

摘要

#$%^&*AU2020100705A420200618.pdf#####ABSTRACT This helmet detection invention is part of the object detection algorithm for safety production monitoring of construction site. It could detect whether construction personnel are wearing safety helmets or not, and identify the level of construction personnel by the color of the helmet. This detection model is based on YOLOv3 network, aiming to improve the detection speed and distinguish the target and background better. More specifically, our model has been improved its algorithms to maintain the efficiency and accuracy of object detection and enhance the recognition ability of small objects. In essence, YOLOv3 is a deep convolution neural network with regression function. The main purpose of YOLOv3 is to predict six parameters from the Bounding Box through multiple layers of Darknet-53: the center coordinates of (x, y), length, width, confidence and the conditional class probabilities. And then uses the Darknet lightweight framework to process images at a faster speed. In this invention, we use transfer learning skill to put pre-trained weights configuration of the helmet to learn our specific helmet training dataset. Through this method, it shows that it has higher detection quality and less detection error in the detection task of high-quality objects. 1512 I 512 128 128 DBL RES DBL RES 9L RES DBL Figure 3.1 e Figure 3.2 2