Design And Application Of Intelligent Terminal For Driving Behavior Detection

The dangerous driving behavior in the field of automobile traffic,especially the fatigue or distracted driving in the "two passengers and one danger" vehicle,deeply endangers people’s life and property safety.Therefore,an important research topic in the field of intelligent transportation is how to reduce the occurrence of such behavior as much as possible through supervision,and intelligent monitoring and alarm device is an important carrier.However,the current monitoring terminal in the market has different forms and methods,and the chip types used include FPGA,GPU,brain-like chip and NPU,etc.Most of them are difficult to balance the contradiction between the performance,power consumption and cost of their hardware devices,which is one of the reasons that the driving behavior monitoring device based on visual method is difficult to land.To solve this problem,this paper starts from Rockchip RV11xx chip platform,which has excellent performance in all aspects,studies and explores the deployment and application methods of advanced deep learning model on this platform,and builds a complete embedded platform driving behavior monitoring and warning system by combining the camera,horn and other hardware.In this paper,the construction of driving behavior monitoring system is divided into three levels:hardware,algorithm and system software control.Firstly,in terms of hardware,we select the RV11xx chip platform with low cost,high performance and suitable development difficulty,and summarize its hardware and software resources and RKMedia,RKNN and other development frameworks.Finally,we complete the hardware part of the system through modeling,drawing processing and hardware assembly.The work on the algorithm is divided into two levels.The first is based on the evaluation criteria of fatigue and distracted driving,and the design of continuous multi-frame image classification algorithm for video input and hand-held object rule to judge fatigue and distraction.Secondly,when the platform is deployed around the Yolov5s model,Focus and large-size Maxpool operators that cannot be supported are replaced.The asymmetric Uint-8-bit plastic quantization method adopted in the deployment model and the image adaptive scaling and filling algorithm to ensure the image scaling ratio remains unchanged are studied.And a multithreading mechanism to accelerate the end-side target detection.Finally,the model achieves 86.6%mAP detection accuracy and 14.5fps speed on RV1126.At the level of embedded system software design,this paper divides the system into five functional modules,and introduces modules such as YUV and RGB format mutual transfer,RTSP push stream,RingBuffer and finite state machine.In order to realize real-time operation,a four-thread concurrent programp is developed.The system built in this paper has the functions of low power trigger and operation,dangerous driving behavior detection and warning,and result visualization.The actual test results of the system prove that it meets the design requirements and realizes the goal of using RV11xx platform to develop a driving behavior monitoring system with high precision,real-time video input processing and early warning.