State Recognition Of Driver Handheld Call Based On Convolutional Neural Network And Implementation Of Embedded System

One of the main causes of traffic accidents is that driver’s hand-held calls during driving.Therefore,the convolutional neural network based on driver handheld call detection and early warning is of great significance to reduce traffic accidents.Object detection has always been a hot topic in the field of computer vision,but the mainstream algorithm has poor detection effect on small target objects,and the model parameters are huge,so it is difficult to apply to mobile devices with limited resources.In view of the above problems,this paper studies the lightweight and application of object detection network.The main tasks are as follows:First,on the basis of the Mobile Net V2-SSDLite object detection network,a new lightweight object detection network LMS-DN is designed by integrating the 5×5depth convolution kernel and extracting the features of two special convolution layers to detect targets.In this paper,9800 driver pictures in the real driving environment are collected,and the Safe＿Imgs data set is made by using the Label Img image annotation tool.Compared with original object detection networks on KITTI,VOC2007+2012and Safe＿Imgs data sets,the experimental results show that the LMS-DN’s m AP on KITTI data set is improved by 2.6% and 3.6% compared with the original model on VOC0712 data set,and the accuracy of LMS-DN on Safe＿Imgs data set can reach86.2%,effectively improve the detection ability for small target objects.The object detection network based on SSD has a very fast detection speed,which has certain advantages in the application of real-time scenes.However,in the case of complex background information,the detection accuracy needs to be further improved.Second,based on the high-precision model YOLOv3,a new lightweight object detection network YOLOv3-X is proposed.According to the network characteristics of YOLOv3,the network is pruned appropriately to simplify the model size,and the detection speed is improved while the detection accuracy is maintained.A scale prediction branch was added to extract the multi-scale prediction feature map for fusion,and the 3×3 convolution of the upper sampling structure was split into 3×1 and 1×3spatial decomposition convolution,which effectively reduced the number of parameters generated by detection.The experimental results show that the accuracy of YOLOv3-X in the Safety＿helmet data set increases by 0.89% while the FPS increases by 18;while the accuracy of Safe＿Imgs data set reaches 93.16%,the number of model parameters decreases by 17 MB,which further improves the detection effect of small target objects.Compared with LMS-DN network,a small amount of detection speed is used in exchange for higher identification accuracy.Thirdly,the effective lightweight object detection network LMS-DN and YOLOv3-X network are applied to the NVIDIA Jetson TX2 development board,and the algorithm is optimized by using the Tensor RT framework.In the embedded terminal,the detection speed and energy consumption of different models on the embedded platform were compared.Combined with the actual vehicle conditions,the detection effect of the above algorithm under different lighting and obstacle shielding conditions was comprehensively tested.Finally,the algorithm was applied to the onboard equipment and video monitoring platform.The experimental results show that the proposed algorithm can well balance the detection accuracy and speed,meet the requirements of real-time detection,and identify the driver’s handheld call.