Automatic Driving Visual Perception Oriented Neural Network Lightweight Study

With the continuous development of autonomous driving technology and deep neural networks,the combination between the two is getting closer.Statistically,about 80%of selfdriving vehicles equip visual sensors for obtaining robust information from the driving environment.These image information obtained by visual sensor can be further processed by depth neural network model.In order to ensure the safety of self-driving vehicles,the automatic driving technology often has higher requirements for real-time performance,that is,the selfdriving vehicle needs to make decisions on the external environment in a short time.In addition,the limitation of computational ability and storage resource are required to be concerned.In order to solve the complex problems encountered in practical application,the current deep neural network model tends to be larger and deeper,and that will require more and more storage space and computing resources.In order to apply the deep neural network model to the autonomous driving platform efficiently,it is necessary to lighten the deep neural network model.This thesis mainly uses the network model pruning and weight parameter quantization to lightweight the network model.Firstly,in terms of network model pruning,this thesis proposes a hybrid multi-granular pruning method to pruning the network model.The network model is pruned by combining channel pruning and weight pruning.In terms of channel pruning,this thesis proposes a layerby-layer channel pruning method based on learnable parameters to pruning the output channels of the network model.By using layer-by-layer channel pruning,the pruned network model has the same ability to extract features as the original network model.After channel pruning,in order to further improve the compression ratio of the network model,we use a probabilitybased weight pruning method to prune the network model,and guide us by using the changes in network model weights.Through the combination of above tow strategies,the pruned network model perform faster with less number of parameters.Finally,through the Jetson Xavier automatic driving platform experiment,the multi-granular hybrid pruning operation proposed in this thesis can achieve good results in speeding up the network model and reducing the network model size.Secondly,in order to use the lower bit to represent the network model,this thesis employs a novel method to measure the weights in a certain network model to achieve the weight lighting purpose.In order to solve the problem that the accuracy of the current low bit method is seriously reduced after quantization,we propose a dynamic quantization method based on the competitive mode.By using the competitive mode of batch by batch quantization for the weight parameters of the network model,the weight parameters of the final network model are quantized.Experiments on the Jetson Xavier automatic driving platform show that when we use the lower bit to quantify our network model.The accuracy of the network model and the original network model can be kept the same.