Design And Implementation Of Local Hardware Accelerator For Lane Line Recognition Based On Deep Learning

In recent years,with the popularity of automobiles on a global scale,more and more conveniences have been brought to people’s lives.However,there has been an increasing number of traffic accidents.Therefore,the research of Advanced Driver Assistance Systems is imminent.Among them,the vision-based lane line recognition technology as an important part of the driving assistance system has become the focus of current research.However,the lane line recognition model based on deep learning proposed generally has the characteristics of large calculation amount and parameter scale,and it is difficult to apply to application such as limited computing resources.Therefore,this thesis designs and implements a local hardware accelerator for lane line recognition based on deep learning from two aspects: a lightweight lane line recognition model based on deep learning and its algorithm hardwareization.In terms of the design of the lane line recognition algorithm model,this thesis designs a relatively lightweight lane line recognition model by introducing the empirical information of the lane line based on the current research status of lane line recognition at home and abroad,which greatly reduces the parameter scale and calculation amount of recognition model at the expense of limited accuracy.In addition,in order to more effectively use hardware to accelerate the calculation of the network model,this thesis adopts the incremental network quantification method for the network model.Quantize the weight parameter of floating-point into a power of 2,and use 4-bit binary bits to encode the quantized result,then,the large number of floating-point numbers operations when the network model is inferred forward are transformed into fixed-point addition and shift operations,which is more conducive to the hardwareization of the network model.Then,the hardware accelerator is realized based on Xilinx PYNQ development board.Finally,based on the Tusimple lane line data set,this thesis conducts accuracy test,speed test and application scenario test based on smart car on the local hardware accelerator system for lane line recognition.The test results show that the lane line recognition model and its local hardware accelerator have higher recognition accuracy and faster recognition speed.