Research On Emotion Recognition For Driving Safety In The Internet Of Vehicles Scene

Driving safety-oriented intelligent detection technology has become the main development direction of a new round of technological reform in the field of Internet of vehicles.As an important component of driving safety oriented,driver facial detection and emotion recognition systems have high driving value for improving safe driving.Therefore,the study of efficient and accurate facial detection and emotion recognition systems for drivers has become a mainstream research direction in the field of the Internet of Vehicles.Although domestic and foreign researchers have made a lot of research progress in this direction,there are still many pain points to be solved.In view of various problems,this thesis mainly makes the following aspects of work:Aiming at the problem of low accuracy of emotion recognition caused by redundancy of nonkey feature information,this thesis proposes an improved MTCNN face detection model.Firstly,the input image is downsampled according to the scaling factor coefficient to build the image pyramid.Secondly,the depth separable convolution is used to replace the ordinary convolution in MTCNN,which makes the network more lightweight and is conducive to the model deployment of the Internet of vehicles scenario.Then,the improved multi-scale feature fusion module is used to fuse the multilevel features,and the feature information of the shallow network and the deep network is transmitted interoperatively to increase the receptive field of the model.Finally,the weight of the facial feature point localization task is optimized,so that the network can locate the feature points more accurately,as to lay a stable foundation for the subsequent driver emotion recognition task.The ablation experiment and multi-model comparison experiment prove that the face detection algorithm proposed in this chapter has better detection effect and better robustness.Aiming at the problems of low accuracy caused of emotion recognition caused by redundancy of non-key feature information,this thesis proposes a deep separable network and region of interest emotion recognition model DSRIN.This model is based on the position coordinates of key facial feature points,performs facial alignment and feature stitching of regions of interest,and eliminates non-key facial feature information.The backbone network was constructed using depth-separable convolution,batch normalization layer and activation function layer,and the control feature dimension of bottleneck residual module was fused,and h-swish was selected as activation function.The experimental results show that compared with the mainstream emotion recognition models,the proposed model is lighter and performs better in accuracy,recall rate and F1 indexAiming at the problems of a large amount of data and high computation delay in the task processing mode of the emotion recognition system,this thesis proposes a driver emotion recognition system based on end-edge-cloud collaboration EEC-Net.Firstly,the system deploys the image compression algorithm based on the key frames of the video stream on the side of the vehicle terminal.The facial video stream is obtained by the acquisition device,and the keyframes in the video stream are extracted by the inter-frame difference algorithm for lossless compression and reconstruction.The reconstructed data is delivered lightweight to the model in the edge computing node for negative emotion monitoring,and the cloud implements online model training,so as to dynamically adjust the operating parameters of the edge model.Simulation results show that EEC-Net can effectively extract video stream keyframes and compress data losslessly,realizing lightweight data delivery.In addition,compared with the traditional centralized processing mode,EEC-Net can greatly reduce the consumption of computing resources and storage resources in the whole system on the premise of ensuring the stability of emotion recognition tasks,and further ensuring the timeliness of the system.