Design And Implementation Of Lane Change Safety Reminding And Driver Status Monitoring Algorithms On Embedded AI Platform

Traffic accidents threaten people’s life and property safety all the time,especially in recent years,the loss caused by traffic accidents cannot be ignored.To avoid the occurrence of traffic accidents due to the complex driving environment and abnormal driving conditions,the realization of assisted driving based on computer vision technology has become a hot research issue,and the development of safe and reliable vehicle-mounted auxiliary lane changing system and driver status real-time monitoring system has important practical value.This thesis mainly studies the algorithm design of vehicle lane change safety reminding in real-time scenarios and the design of driver status monitoring algorithms as well as their implementation and application on the embedded AI chip.To avoid the problem of complex labelled data are required for training vehicle lane change safety reminding algorithms in real-time scenarios,this thesis proposes an end-toend "lane change related vehicle" detection scheme that does not need to label and recognize lane lines,and by improving the Mobile Netv2+SSD detection framework to achieve accurate and rapid detection on the embedded AI chip.Firstly,the vehicle label is marked with lane line constraints when the data set is made,and only the vehicles related to the lane change are marked,and the deep convolutional neural network is used to learn the high-level semantic position information of the vehicle relative to the lane line,to avoid complicated lane markings.Then,all the corresponding detection heads are designed in a deeply separable form,so as to ensure the speed of model reasoning and the real-time performance of the system.Finally,in response to vehicle missed detection and background false detection,a sparse self-attention mechanism is added to the corresponding feature layer,which improves the accuracy of detection while ensuring efficiency,and and the whole system reaches the standard of practicability.Regarding the real-time monitoring system for driver status has the problem of complex implementation process,this thesis designs a new,simple and effective technical solution for the driver’s "integrated face + local joint parts recognition".The first step is to perform regional detection of the key parts of the human body in the input image,including the face,eyes,mouth and hands.The second step is to classify the status of the detected area to determine the specific status of the human body(normal,distracted,smoking,talking on the phone,tired).The status of smoking and calling is mainly distinguished according to the status of the hand area.Considering that the eyes and mouth are small targets relative to the input picture,the strategy of stitching the zooming in the face area and zooming in the area of the eyes and mouth is used for classification to distinguish distraction and fatigue.Finally,the RK1808 embedded AI platform is selected according to system requirements,and the model is deployed based on the ARM board.This thesis realizes realtime and accurate lane change safety reminding and driver status monitoring through the proposed lane change-related vehicle detection and overall facial joint recognition schemes and algorithms,and it has been integrated into the on-board equipment and put into use.