Automatic Driving Environment Perception Oriented Neural Network Model Compression Study

In recent years,deep learning neural network technology has been continuously developed,which plays a great role in dealing with image information related problems.The emergence of autonomous driving technology is very helpful to the improvement of traffic safety.In the autonomous driving technology,the perception of the surrounding environment is the first step of the whole process.Only accurate and timely information about the surrounding environment can guarantee a good guiding role for the subsequent decision-making process.The environment perception link mainly deals with the image information of the surrounding environment extracted by the camera.In order to process complex image information,the existing neural network model usually has a lot of parameter parameters,and the requirements for computing resources and storage space will be relatively large.Many models that can run well on the server platform are directly transplanted to the automatic driving platform.The running speed will be greatly limited,mainly because the computing power and storage capacity of the automatic driving platform are relatively limited,and in the traffic scenario of automatic driving,real-time performance is an essential requirement.Therefore,in order to ensure that the deep neural network model can run on the automatic driving platform in real time and accurately,it is necessary to compress the existing deep neural network model.This article considers two aspects of the structure of the deep neural model and the preservation of the model weight parameters.Compress the deep neural network model.First,for the compression research of the neural network model structure,this paper proposes a channel pruning method based on sensitivity and scale parameters.First,the sensitivity of each network layer to be pruned is calculated according to the scale parameters of the channel layers of the network obtained by training Then,determine the pruning rate of each layer according to the sensitivity.Finally,pruning the network layer by layer,the model after pruning is simply retrained to restore accuracy.Finally,through experimental verification on the automatic driving platform Jetson Xavier,the pruning method in this paper can improve the running speed and compression rate of the model while ensuring a small loss of accuracy.Second,for the study of the parameter storage method of the neural network model,this paper proposes a dynamic grouping low-bit weight parameter quantization method to solve the existing low-bit quantization method after the weight parameter is quantized.problem.By dynamically grouping the quantized weight parameters and unquantized weight parameters,the quantized model parameter distribution is close to the original model,and through experimental verification on the automatic driving platform Jetson Xavier,it is concluded that when using 5bit After quantifying the original network model,it can compress the model’s volume well,and ensure that the accuracy of the model is close to the original model.