Roadbed Traffic Sign Detection,Identification And Driving Decision Support In Urban Environment

Accurate identification of traffic information is an important guarantee for safe driving of driverless vehicle on the road,including information of pedestrian,other information of vehicle in the road,information of traffic sign such as speed limit and turning,traffic signal and so on.This basic information can be accurately identified in order to provide decision-making basis for subsequent operations such as obstacle avoidance and tracking of vehicles.The image classification based on vision is to classify the identified objects.In the unmanned driving system,it is to judge the traffic information,such as turning left,going straight or lifting the speed limit sign,and make further planning for driving according to the judgement results.In the city environment,keeping driving on the lane line has important significance to the whole traffic system,in recent years,methods based on the deep learning performed very well in the lane detection task.its powerful representation ability could finish various tasks even the information of target has been partially occluded.A fter the lane line is detected,if there is an abnormal deviation of the vehicle,a deviation warning can be issued.This is the deviation warning.During the driving of the vehicle,some unexpected situations often occur,such as traffic signs are blocked or the correctness of the detection results cannot he verified.Continuous tracking of the target can effectively solve such problems,and at the same time,the complexity of the road,such as Special road sections such as tunnel entrances or ramp junctions are high-accident areas.This paper analyzes the decision-making behavior of vehicles by using a high-precision map-based finite state machine method.The research work of this paper is as follows:1)In view of the fact that data labels are often missing in real scenarios,a semi-supervised autoencoder is proposed,and the objective function is improved by using the label information,and the robustness of the system classification is improved by the stacking method.This article uses artificial removal of some data labels to simulate the absence of data labels in real scenarios.Pre-training the unlabeled data to improve the model's convergence speed during training,and then deepening the depth of the network by stacking to extract more abstract hidden layer features.The objective function is trained using the gradient descent method.For the output of the autoencoder,we use the softmax function for smoothing to get the final result.The algorithm is tested on the traffic sign datasets Belgium TS dataset and GTSRB dataset.At the same time,experiments on SVHN and CIFAR-10 datasets to verify the effectiveness of our proposed algorithm,which proves the effectiveness of the improved algorithm.2)In view of the fact that lane lines are often occluded in real scenarios,this paper proposes an improvement solution based on the spatial convolutional neural network model.When the spatial convolutional neural network transmits feature information,features are transmitted in four directions:up,down,front,back,and in order to extract the spatial feature information of the lane line.For the extraction of local information,a good structure is the attention mechanism,which can automatically extract local features and filter out useless noise such as background information.Attention gates can be used to extract local information for the propagation layer in each direction.This feedforward structure can combine the overall information and local information through additive operations,sequentially,improving the model's ability to extract features.The amount of calculation is an important reflection of the complexity of the model,and such an attention mechanism does not increase the amount of calculation too much,ensuring the calculation efficiency of the model.The model was tested on the new lane line detection CULane dataset,and the results show the effectiveness of the model.3)Aiming at the problems such as occlusion,frame loss,and failure to verify the correctness of the recognition results during the process of the detection of traffic signs,this article uses multi-target tracking technology to accurately track traffic targets to ensure continuous recognition during driving.For the situation that lane lanes on the road may not exist,and in some important intersections,such as tunnel entrances,ramp junctions and other accident-prone areas,it plays a vital role in vehicle behavior decision analysis.At the same time as the multi-attribute decision-making method,a finite state machine based on a high-precision map is also proposed,which can avoid some dangerous operations a little,thereby ensuring the safe driving of the vehicle.At the same time,this paper analyzes the vehicle decision-making process with an example of multi-attribute decision analysis.