Driving fatigue detection method and system combining pseudo 3D convolutional neural network and attention mechanism

Abstract

The invention discloses a driving fatigue detection method combining a pseudo 3D convolutional neural network and an attention mechanism. Including: 1) Extract and process the video frame sequence from the driving video; 2) Use the pseudo 3D convolution module to learn the spatiotemporal features; 3) Build the P3D‑Attention module, and use the attention mechanism to pay attention to the channels and feature maps; 4) A 2D global average pooling layer is used instead of a 3D global average pooling layer to obtain more expressive features, and a softmax classification layer is used for classification. The present invention can analyze yawning, blinking, and head feature movement, and can well distinguish yawning behavior from speaking behavior; effectively distinguish the three states of vigilance, low vigilance, and drowsiness, so as to improve fatigue driving behavior. predictive performance.

Description

technical field [0001] The invention relates to the technical field of intelligent video analysis, in particular to a driving fatigue detection method and system combining a pseudo 3D convolutional neural network and an attention mechanism. Background technique [0002] Fatigue driving is one of the main causes of traffic accidents. Drivers who are in a state of fatigue often feel drowsy, temporarily lose consciousness, reduce alertness and ability to deal with abnormal events, and slow down the reaction time to traffic control and danger, thus leading to accidents. The American Automobile Association reports that driver fatigue accounts for the largest percentage of road accidents, with 7 percent of all accidents and 21 percent of fatal crashes caused by tired drivers. Existing technologies for detecting fatigue driving behavior can be divided into three categories: based on physiological parameters, based on vehicle behavior and based on facial feature analysis. [0003]...

AI Technical Summary

Problems solved by technology

However, due to the use of GoogleNet and Inception-v3 models for spatial feature extraction, the prediction model has a huge amount of parameters and contains a large amount of redundant spatial data. The convolutional spatial features are converted into one-dimensional vectors and input into the time series model, which does not take into account the spatial Correlation, and did not remove background and noise interference to the recognition, resulting in spatio-temporal features are not well integrated

Images