Research And Design Of A Driving Behavior Detection System Based On Deep Learning

Nowadays,the automobile industry has developed unprecedentedly and the automobile manufacturing technology has continuously improved.Although the automotive industry provides security for the safe and reliable travel of modern vehicles,there are still many traffic accidents caused by dangerous driving behaviors of drivers.Such as fatigue driving behavior and distracted driving behavior.Dangerous behavior inevitably reduces the driver’s ability to control motor vehicles.Extremely liable to cause traffic accidents.Therefore,this article analyzes and studies the causes of traffic accidents.Research methods for identifying fatigue and distracted driving based on deep learning related technologies.And develop relevant detection systems.The system reminds the driver to drive safely.It is of great significance to reduce traffic accidents and build a safe China.The main research work of this study can be summarized as follows.Ⅰ.Design of a driving behavior detection system scheme and analysis of related technologies.The system consists of hardware and software.The hardware part includes data acquisition module,microprocessor and display module.The software part includes fatigue driving detection algorithm and distracted driving detection algorithm.Then,the relevant technologies involved in the driving behavior detection system are analyzed in depth.It includes convolution neural network theory for image processing,short-term and short-term memory network theory for learning sequence features,and graph convolution network theory for graph depth learning.It lays a foundation for the design of relevant algorithms of the following driving behavior detection system.Ⅱ.Research and design of driver fatigue detection algorithm.Addressing the coexistence of long-term and short-term dependencies in fatigue detection features.Propose a fatigue detection algorithm based on improved ResNet-50 and Multiscale Time Skip LSTM.Firstly,MTCNN is used for face detection and key point location.Design clipping algorithm to extract head,eye and mouth images.Secondly,the extracted image is input into the improved ResNet-50 network and facial spatial features are extracted.The improved ResNet-50 integrates shallow,middle and deep features by reserving effective layers and multi-scale.Finally,input MTS-LSTM to extract temporal features.MTS-LSTM fuses the sequence features of different time skips.It makes the network better deal with the dependence of different fatigue behaviors on the spatial features of long and short periods.In the experiment part,we designed the space feature extraction contrast experiment,the temporal features extraction contrast experiment,the face occlusion experiment and the accuracy contrast experiment.The accuracy,loss value and model parameters are used to evaluate the algorithm performance.The accuracy of the fatigue detection algorithm proposed in this study is 89.72%in YawDD dataset and 90.87%in NTHU-DDD dataset.Ⅲ.Research and design of driver distraction detection algorithm.Many action recognition algorithms based on graph convolutional networks currently face problems such as large network scale,excessive redundant information,and difficulty in training.Propose Two Stream Hybrid RefinementCorrection Graph Convolutional Network(2sHRC-GCN).First,AlphaPose is used to detect human key points and estimate human posture.The algorithm obtains video human joint data and edge data.Secondly,the joint sequence and edge sequence data are input into 2sHRC-GCN for feature extraction.HRCGCN transmits the sequence information of the correction network to the refinement network through lateral connections.In addition,the output features of refinement network and correction network are better preserved by adaptive feature fusion.In the experiment part,the distraction behavior detection experiment,2sHRC-GCN ablation experiment,HRC-GCN configuration experiment and accuracy comparison experiment were designed.Precision,accuracy and recall are mainly used to evaluate the algorithm performance.The accuracy of the distracted driving detection algorithm proposed in this study is 93.48%on the SUST-DBDD and EBDD data sets.Ⅳ.Design and Implementation of a Driving Behavior Detection SystemDesign and implement a driving behavior detection system based on driver fatigue detection algorithm and driver distraction detection algorithm.Firstly,research the design concept of the detection system and establish design principles.And study the overall architecture of the driver behavior detection system.Design the hardware and software parts of the system based on the information perception layer,data transmission layer,and data processing layer of the detection system.Capture the driver’s behavior inside the vehicle through an onboard camera.Then upload the data to the server.Evaluate driving behavior through fatigue detection algorithms and distraction detection algorithms.Finally,complete the implementation of the detection system and nighttime detection.And test the functionality of the behavior detection module.Based on deep learning related technologies,the driving behavior detection system is studied through several parts:design and analysis of driving behavior detection system scheme,research and design of driver fatigue detection algorithm,research and design of driver distraction detection algorithm,implementation of driving behavior detection system and visual display interface design.Technologies such as multi-scale feature extraction and refinement correction framework are applied in the system.In addition,algorithm structures such as multi scale time hopping and attention adaptive feature fusion are also proposed in the research.These measures reduce the amount of parameters generated during network detection and improve the accuracy of behavior classification.The sensitivity and generalization of the detection system have also been greatly improved.The driving behavior detection system in this article can provide some reference for behavior detection in various fields.