Research And Application Of Traffic Sign Detection Based On Deep Learning

Traffic sign detection technology is a hot and difficult issue in the field of target detection.In actual scenes,streetscapes are complex and diverse,traffic signs account for a very small proportion of the entire picture.When feature extraction is performed,the characteristics of traffic signs are often affected by the surroundings and small-sized targets,such as billboards and other objects.Therefore,the detection effect is poor in reality.In addition,traffic sign detection systems are usually mounted on mobile platforms such as smart cars,these systems need to quickly and accurately identify front signs within a very low time delay.However,it is difficult for existing methods to achieve a balance between detection accuracy and speed.In response to the above problems,this thesis proposes a traffic sign detection method based on improved YOLOv3.The specific work mainly includes the following aspects:Firstly,a traffic sign detection method based on feature fusion is proposed.After maximum pooling,mean pooling and convolution operations,the shallow features are spliced and fused in different forms with the deep features.Then the fused features are used for detection to solve the problem of serious information loss and dilution of semantic features of small-size traffic sign in the deep feature map.Secondly,based on the fusion and multiplexing of fine-grained feature maps,the importance of different feature channels in the detection task is further analyzed after feature splicing.In order to distinguish the fine-grained features output by the shallow position of the network(location information is more accurate)and the coarse-grained features output by the deep position of the network(more semantic information)in the target detection task,an attention mechanism is introduced in the detection part of the network.Thirdly,this thesis builds a traffic sign detection system based on Atlas 200 DK equipped with Ascend 310 AI processor.After training the model using Tensorflow deep learning framework,the original model is converted to an OM model and deployed on Atlas 200 DK.The object system module calls an external camera to realize the real-time detection of traffic signs.