Faster Detection Method of Driver Smoking Based on Decomposed YOLOv5

Abstract

Abstract Aiming at the problem of large computational resource consumption in the existing object detection network for driver smoking detection, a decomposing YOLOv5 network (Dec-YOLOv5) is proposed for optimization. This method uses singular value decomposition (SVD) to split the standard convolution in the pre-trained YOLOv5 network into two simpler convolution operations to reduce the computational cost. The optimized network generated after decomposition does not need to be retrained, which can reduce the number of parameters and calculations while maintaining the detection accuracy of the pre-trained model. The experimental results show that the detection time of Dec-YOLOv5 is only 80% of the original YOLOv5 when the overall detection accuracy reaches 93.5%. At the same time, compared with the current mainstream object detection model, the Dec-YOLOv5 network has a better ability to express the characteristics of the driver’s cigarettes, and the detection accuracy is higher.