Research On Road Traffic Sign Recognition Technology In Autonomous Driving Scene

With the rapid development of economy and science and technology,autonomous driving technology has become a hot topic in recent years.Traffic sign detection is one of the core technologies of autonomous driving.Traffic sign identification can not only improve the road capacity and predict the road condition,but also reduce traffic accidents and save energy.Therefore,the fast and reliable traffic sign recognition system has become a crucial part of the automatic driving system.However,the road scene in the real environment is complex and changeable,and the existing traffic sign detection and recognition methods still have room for improvement in real-time and accuracy.Since Alex Net Image Net Challenge: ILSVRC achieved a record of 2012 since the degree of image classification,convolution neural network has been ubiquitous in computer vision,build deeper and more convolution neural network is the main trend of visual identification tasks,however,in many practical applications,such as automatic driving,intelligent monitoring,real-time computing resources limited platform is needed in the identification or testing tasks,especially edge calculation is put forward and development of the made this demand more urgent.Based on target detection based on deep learning mainstream architecture and related subproblems,on the basis of full investigation,analysis of the existing technology in theory,algorithms and application problems in the research and improvement,and made the corresponding research work is aimed at building a lightweight,efficient target detection network,mainly includes the following aspects:1.Improved depture-separable convolution unit: the idea of grouping convolution is introduced to replace the 1x1 convolution in the standard depture-separable convolution with grouping convolution.By making each convolution operate only on the corresponding input channel group,the number of parameters in the convolution layer is reduced.At the same time,Leaky Re LU was used as the activation function after grouping convolution.The comparison test on the MNIST data set showed that Leaky Re LU’s performance was better than Re LU,especially when the network was deep,the gap was more obvious.2.Channel mixing strategy: when multiple groups are added together,there is a side effect: the output of a channel only comes from a small part of the input channel.Obviously,the output of a group is only related to the input within the group.This property blocks the flow of information between channel groups and weakens the presentation.Using the strategy of channel mixing,the convolution is allowed to get input data from different groups,so that the input and output are related.3.Improved fast connection: this paper is a feature extraction network based on the identical fast connection unit of Res Net.The difference is that this paper adds a 3×3 average pool on the fast path.At the same time,the channel cascade is used to replace the element cascade,which facilitates the amplification of channel dimension and makes the extra calculation cost small.4.Construction of road traffic sign detection framework: the improved depth-separable convolution unit is applied to the standard convolution in fasters-rcnn prediction layer,and the feature extraction network constructed in this paper is used to replace the original feature extraction network to form a new target detection framework.Based on the above research,this paper designed a lightweight target detection network,and compared the network performance on Image Net2012 and its own traffic sign data set.Experimental results show that the proposed network model has better recognition accuracy and speed than the current mainstream network model,especially in the case of extremely limited computing resources.