Blinking frequency and sight line estimation method based on eye region generation network

Abstract

The invention discloses a blinking frequency and sight line estimation method based on an eye region generation network. The method can be used for carrying out early warning on fatigue driving of a driver. Particularly, the method comprises the steps of: firstly, applying machine deep learning to extract an eye feature region of the driver, which is shot by a camera; carrying out driver eye simulation on the feature region by an ellipse fitting method; and determining a blinking frequency and pupil positions by combining perclos. The estimation method gives out the blinking frequency and a sight line direction of the driver and provides supporting bases for a driver fatigue monitoring system and driver dynamic behavior analysis.

Description

technical field [0001] The invention belongs to the field of automobile safety and relates to a method for estimating the driver's blinking frequency and line of sight, and the method can be used for early warning of driver's fatigue driving. Background technique [0002] At present, traffic accidents caused by fatigue driving still occupy the main component. Therefore, the successful early warning of drivers under driving fatigue has become a problem that people pay more and more attention to. This has also led most researchers to propose many effective auxiliary measures for this problem. The development of image processing technology and the extraction of facial features have provided a theoretical basis for the research on driver fatigue warning methods, but they are still immature. period. E.g: [0003] Since the target may appear anywhere in the image, and the size and aspect ratio of the target are also uncertain. In the process of traditional target detection, the...

AI Technical Summary

Problems solved by technology

In the process of traditional target detection, the sliding window strategy is used to traverse the entire image, and different scales and different lengths and widths need to be set, which makes the complexity too high, produces too many redundant windows, and takes a long time

This seriously affects subsequent feature extraction

In the deep learning target detection algorithm of Region Proposal (candidate area), the extraction of candidate areas has been greatly improved, but it is still not close to the real range of the target. The R-CNN training under development is divided into multiple stages. It is cumbersome, time-consuming to train, and takes up a lot of disk space. It takes 47s to process an image using the GPU, VGG16 model, and the processing speed is slow. However, the improved versions of SPP-NET, Fast R-CNN, and Faster R-CNN, although in terms of processing speed and The processing steps are simplified, but the extraction of candidate regions takes a lot of time and cannot meet the real-time requirements

Images