Deck officer drowsiness detection based on Improved GhostNet-SSD and grey correlation analysis

Abstract

Deck officer drowsy driving has become a major cause of shipping accidents, while traditional drowsiness detection methods have struggled to cope with the complex detection environment of a ship's cockpit. In this paper, we propose a deck officer drowsiness detection method based on the improved GhostNet-SSD.A multi-scale feature extraction network is constructed on the basis of lightweight GhostNet to generate redundant feature maps by depthwise separable convolutions. Extracting features of small targets at multiple scales facilitates that the model can be deployed on shipboard low-performance devices and enhances the detection accuracy of the model at the same time. In the inference stage, an improved soft-NMS algorithm is proposed to optimize the process of removing overlapping prior bounding box, reduce the miss rate of overlapping targets, and boost the detection speed of the model.