Objective Due to the current problem that driver fatigue detection cannot simultaneously balance detection speed and accuracy this paper proposed a fatigue driving detection algorithm based on an improved YOLOv7 model. Methods Firstly to improve the model?? s convergence speed and enhance its detection performance the traditional convolutional layer was replaced with a depthwise over-parameterized convolutional layer which accelerated the fitting process by adding learnable parameters. Secondly in scenes with occluded targets traditional downsampling processes can lead to significant feature loss. To improve the accuracy of occluded target detection the algorithm introduced a Depthwise Separable Convolution DS-Conv module based on improved Squeeze-and-Excitation Attention. Thirdly to enhance the model?? s ability to detect small targets an MSS attention module with multi-scale feature extraction was added to the feature extraction layer which can capture the details and contextual information of targets at different scales. Finally fatigue driving determination was performed on the detected faces according to the PERCLOS criterion. Results The experimental results showed that the improved algorithm achieved accuracies of 96. 0% 94. 6% and 88. 1% on the Easy Medium and Hard subsets of the WIDER FACE dataset respectively. Conclusion The improved algorithm with its simple structure and small number of parameters is conducive to real-time face target detection and is suitable for deployment in resource-limited environments such as in-vehicle systems effectively ensuring driver safety.