Research On Driver Hand Detection And Driving Behavior Analysis

Hands are important organs for humans to interact with the outside world.Basic behaviors such as eating and drinking in people’s daily life are inseparable from the participation of hands.Hand plays an important role in the interaction between driver and car in driving environment.In recent years,hand detection and behavior analysis based on hand detection have become one of the research hotspots in the field of computer vision.In driving assistance,the recognition of dangerous driving behaviors and the detection of driving distractions can be realized through the detection of the driver hand and the analysis of hand behavior,which is of great significance to safe driving.Illumination and occlusion in complex environment bring great challenges to the research of hand detection and behavior analysis.Driver hand detection is a small target detection problem in complex scenes.The analysis of driving behavior is a multi-classification problem.Common implementation methods include device-based hand behavior analysis and deep learning-based hand detection and behavior analysis.Device-based hand behavior analysis relies on hardware devices such as Kinect or data gloves that can accurately identify hand behaviors in a timely and accurate manner through sensors.It can reduce the impact of environmental hand detection,which is of great significance to the development of human-computer interaction.However,the requirements of the scene and equipment make its application scene limited.The hand detection and behavior analysis based on deep learning rely on the development of deep learning,and the learning effect of convolutional neural network on features is continuously improved.At the same time,the accuracy of hand detection and behavior analysis was also significantly improved.This paper is based on deep learning to realize the research of driver hand detection and behavior analysis.First,the YOLOv4 target detection algorithm based on the added attention mechanism is proposed to realize the driver hand detection.Second,the CenterNet based on the optimized convolution block is proposed to complete the driver behavior.In addition,the driver’s hand detection data set and driver behavior data set are collected and constructed for the training and testing of the deep learning model in the algorithm proposed in this paper.The innovation points of this paper are summarized as the following three aspects:(1)In our research,we collected and constructed the driver hand detection data set and the driving behavior data set.A driving recorder was employed to collect the driving videos of the drivers and the driving videos of the five behaviors of smoking,drinking,eating,calling,and watching the mobile phone.The collected videos were screened and processed into pictures,and then the pictures were labeled.The driver hand detection data set and driver driving behavior data set suitable for network training and testing were constructed.(2)Aiming at solving the driver hand detection in complex driving environments,the target detection algorithm YOLOv4 was adopted.By adding the attention mechanism network,the YOLOv4 driver hand detection algorithm based on the SE module of the attention mechanism was proposed.This paper optimized the backbone feature extraction network CSPDarknet of YOLOv4.Add the attention module Squeeze-and-Excitation Network between each convolutional layer of CSPDarknet.This could strengthen the training and learning of the characteristics of the CSPDarknet backbone network,and improve the accuracy of the hand detection.(3)Focal Loss based CenterNet algorithm was adopted for driving behavior analysis,and the CenterNet algorithm of block was optimized by adding channel attention mechanism.The backbone feature extraction network of the CenterNet algorithm was Res Net50.The block of Res Net50 included the Conv block for downsampling and the Identity block for connection.This article added a channel attention mechanism to each block to optimize the network training effect of the residual block.It sped up the flow of features between networks,and improved the effect of driver hand behavior analysis.