Research On Detection And Recognition Algorithm Of Traffic Sign In Natural Environment

The scene is complex in natural environment,how to detect and recognize traffic signs efficiently is one of the research hotspots in computer vision and intelligent transportation system.Traffic sign detection aims to find the areas which contain the traffic sign.Traffic sign recognition aims to determine the specific category of the areas.In the process of driving,automatically detecting the traffic signs in front of drivers and giving the tips can not only reduce the occurrence of traffic accidents,but also can enhance the comfort of driving.It is alao an important part of the driver assistant system and the driverless operation system.To solve this problem,many traditional methods use color-based or shape-based methods to detect traffic signs and recognize traffic signs by using feature extraction and training classifiers.These methods usually perform well on a single data set,but in the face of complex natural scenes,these methods are difficult to have a good performance.On the other hand,some traditional methods treat detection and recognition as two independent stages,and can not do end-to end detection and recognition on the same dataset.The newest methods based on deep learning have been used to detect and recognize traffic signs in images.These methods unify the processes of detection and recognition into a framework and do end-to-end detection and recognition.However,as the speed becomes faster and faster,the traffic signs in the distance in the field of vision need to be detected in advance.However,these traffic signs are relatively small and difficult to be detected by using the common target detection method.In this paper,a multi-task learning model based on deep convolutional neural network is adopted,and we have made improvements on this model,which can improve the detection recall rate and accuracy,and the network can also detect the small traffic signs in the distance.The multi-task learning framework is divided into two parts,the first part is the feature extraction sub-network,for this part,this paper proposes a network with fewer parameters,and feature fusion strategy is also adopted,the improved network is easier to train and can extract more useful information for traffic sign detection.The second part is the multi-task learning sub-network,according to the characteristics of different tasks,detection task and classification task are treated differently,and multi-task learning is performed on different convolution feature maps,which improves the detection accuracy and recall rate.The TT100 K public benchmark is mainly used for experiments.The experimental results verify that the improved multi-task learning model can improve the accuracy and recall rate of traffic sign detection,and it can also detect small traffic signs well.The overall recall rate and accuracy rate on the TT100 K benchmark are 94% and 91% respectively,which are 3% and 3% higher than before.Finally,the generalization ability of the network is analyzed on the other benchmark CTSD,reaching 95.40% recall rate and 97.46% accuracy rate.