Research On Driver’s Emotion Recognition Based On Neural Network

Emotion is the expression of subjective feelings that is not screened and identified,and always affects people’s decisions and behaviors.The sudden emotions of drivers can have a certain impact on safe driving,increasing the likelihood of traffic accidents.Drivers in an angry mood are particularly prone to behaviors such as forced stoppages and aggressive driving.Therefore,it is of great significance to quickly and effectively identify the emotions that affect safe driving,and regulate the driver’s emotions in an appropriate way to maintain a stable range of emotions,in order to improve road traffic safety.However,the complexity of driving environments poses challenges to recognizing driver emotions,including poor accuracy,slow recognition speed,and low robustness.This paper focuses on more effectively identifying driver emotions in realtime.This paper’s specific objectives include:(1)In the problem of slower recognition speed in the face detection of driver’s face,a fast-robbing driver face detection algorithm Ghostnet-SSD.This algorithm leverages the efficient feature extraction capability of Ghost Net and optimizes the backbone feature extraction network of SSD.We also use a transfer learning training strategy to enhance the network’s robustness and generalization ability,resulting in a fast and reliable Ghost Net-SSD model.Our experimental results demonstrate that Ghost NetSSD can effectively detect the driver’s face position information in complex environments,exhibiting strong robustness.Moreover,the algorithm achieves a recognition speed of 83ms/frame on our experimental platform,which is the best among the comparison algorithms.(2)In response to the problem of low accuracy of the algorithm in the emotional recognition of the driver,the emotional recognition model SVGG based on a attention mechanism is proposed.To address the issue of image shaking during driving,the model employs a spatial transformation network to align facial images,enabling more effective feature extraction by the network.We have also improved modules such as the convolution method and activation function to enhance the convergence and reasoning speed of the network.Lastly,the channel attention module ECA-Net is used to reallocate the weight of feature maps between channels,resulting in improved recognition accuracy.Our experimental results indicate that the SVGG model achieves accuracy rates of 96.6% and 72.4% on the KMU-FED and FER-2013 datasets,respectively.Additionally,the model attains a recognition speed of 80ms/frame on our experimental platform,which has good real-time performance.