Improved SSD for crew fatigue driving detection algorithm

Abstract

To solve the problems of low recognition accuracy and slow detection of crew fatigue driving behavior in the cockpit of ships during the process of sailing in and out of the port, the SSD model was studied. By replacing its backbone network and improving the prior frame generation mechanism, the MV-SSD model is proposed. Replace the backbone network VGG16 in the original SSD model with MobileNetV3, reducing the network parameters of the backbone network. Using the K-means algorithm to cluster the real detection boxes in the face area dataset, the prior box allocation mechanism of the SSD model was redesigned, reducing the number of prior box generation by nearly half, and the ERT algorithm in the Dlib library is combined to locate the face key points, and finally, the PERCLOS criterion is used to determine whether the driver is fatigued. Experimental results show that the average accuracy (mAP value) of the MV-SSD model is 7.15% higher than that of the original SSD model, and the detection speed (FPS value) is increased by 98frames/s, which is more suitable for the detection of the crew face area, and the average accuracy of the constructed fatigue detection algorithm for fatigue features is more than 94%.