Research On Traffic Sign Detection Method Based On Deep Learning For Small Embedded Devices

With the improvement of the living standard of Chinese residents,the number of cars in China is also increasing day by day,However,a series of traffic problems arise,such as:road congestion,frequent traffic accidents and so on.It not only brings the driver a bad driving experience,but also affects the driver’s safety greatly.Traffic sign detection system can assist the driver to drive,and can give early warning to the driver during the driving process,so as to improve the driver’s driving experience and reduce the occurrence of traffic accidents.Because of the advantages of low cost and small size,small embedded devices are used in many aspects of life.If the traffic sign detection can be applied to small embedded devices,the cost of the traffic sign detection system can be greatly reduced,which is conducive to the popularization of the traffic sign detection system in the family car and other fields.Therefore,this thesis focuses on the application of traffic sign detection in small embedded devices.Deep learning has the advantages of simple feature extraction and multiple types of recognition,which can be adapted to the traffic sign detection task in complex scenes.Therefore,this thesis chooses the method of deep learning for traffic sign detection.Due to the complex structure of deep learning network,the real-time performance of the deep learning network in low-performance small embedded devices is very low.Therefore,this thesis selects the relatively real-time SSD network and makes a series of improvement measures for the SSD network.The main work of this thesis is as follows:(1)Aiming at the problem of low real-time performance of the SSD network in small embedded devices,this thesis considers embedding Google lightweight network MobileNet as the basic feature extraction network into the SSD network,replacing the more complex VGG16 in the original network.This reduces the network complexity and improves the real-time nature of the network.(2)Aiming at the problem that the accuracy rate decreases after the network complexity is reduced,this thesis designs a branch fusion network to fuse feature layers to better integrate deep and shallow features,thereby improving the network’s accuracy in detecting traffic signs.(3)This thesis analyzes the characteristics of traffic signs,and analyzes that the size of traffic signs is generally below(90 X 90)pixels.In response to this conclusion,the auxiliary convolution network of the SSD network was redesigned and the detection classifier was re-planned.The detection classifier of the original network was simplified,and the detection classifier that was meaningless for traffic sign detection was removed,forming a pair.Four different scale feature layers are used to detect and classify the network.(4)To build a deep learning operating environment for small embedded devices and implant the improved network into it,the experiment shows that compared with the original SSD network,the operating frame rate(FPS)is increased by 3.5 times.