Research On The Recognition And Classification Methods Of Driver's Head And Pose Based On Deep Learning

In today's complex transportation environment,assisted safety driving has become an important method to improve driving safety and is also the research focus and frontier in the field of intelligent transportation.With the rapid development of computer hardware technology and Artificial Intelligence,machine vision technology has again received the great attention of researchers at home and abroad.In this paper,according to the features of driver's head and body posture,combined with Deep Learning,a driving behavior detection method is proposed.The driver's behavior detection function is realized through the driver's head regional positioning and recognition method and the driver's body posture feature extraction and recognition method.The feasibility and effectiveness of the method are verified through relevant experiments.The main work and research results include:(1)Research and propose a real-time driver's head regional localization method based on Yolo algorithm.This method first marks the driver's head area with a tool such as BBox_Label_Tool on a small batch of drivers data and converts it to the area annotation format specified by Yolo.Yolo network structure format is set to adapt to the current project,iterative training to get real-time detection of the driver's head area Yolo network structure parameters.(2)The driver's head feature extraction method based on Deep Convolutional Neural Network is designed.Convolutional neural network structure usually consists of one or several convolutional layers,pooling layers and fully connected layers,whose convolutional layers and the pooling layers can get more accurate image features through multi-layer extraction of image features and combine calculation Description.Transfer Learning refers to the application of related knowledge learned in the past to solve new problems and to get better results while solving new problems and learn faster.This paper makes full use of the feature extraction ability of convolution layers and pooling layers and the advantages of transfer learning,and realizes the classification of driver's different head states based on convolutional network.(3)In order to identify the driver's attitude,a Tiny Pose real-time human body key points detection network structure is proposed,which can extracting the key points locations of the body.Tiny Pose is a simplified version of the real-time human body key detection algorithm Open Pose,which makes Open Pose suitable for the driving environment,and testing speed is faster.(4)Through the BP network(Back Propagation)to achieve the driver's different posture classification and recognition.The driver's body position information obtained from the Tiny Pose network is composed of a training set and a verification set.Design BP network structure and other machine learning models,the driver's gesture classification comparison experiment.Experimental results show that BP network can effectively classify the driver's pose feature.